# Frequency by Infinite Inference > Voice AI built and run in India. Agents from ₹1/min with a volume commitment. Try the sandbox. Website: https://scalefrequency.com Docs: https://docs.miraiminds.co Contact: sneh@miraiminds.co ## Frequency that Indian builders need. Natural conversations. In your language. Voice agents and speech models, built and run in India. ## One conversation. Six or seven products to stitch together. A typical builder connects separate products for telephony, orchestration, noise removal, ASR, LLM and TTS. Getting them connected is only the start. Making them work as one conversation is the hard part. Frequency brings ASR, LLM, TTS and orchestration together. The response diagram compares an illustrative stitched setup with our integrated stack. The illustrative stitched setup takes 1600 ms. This is not a measured competitor benchmark. The Frequency example is 850 ms, including 150 ms LLM and 110 ms TTS. Remaining time covers ASR, VAD and buffering. Actual response times depend on speech, network and load. ## Price for India Every company starts on T5: ₹5/minute, no monthly commitment. T3: ₹3/minute with a ₹50,000 monthly minimum. T1: ₹1/minute with a ₹1,00,000 monthly minimum, optional Mirai telco at ₹0.30/minute and optional insights at ₹0.20/minute. The same models and features, including single-prompt and node-based agents, are available on every tier. T5 and T3 include Mirai phone calling and insights. T1 adds ₹0.30/minute only for Mirai-managed phone calls. BYO telco, WhatsApp and browser calls have no Mirai telco charge on any tier. Provider charges you pay directly are separate. All prices exclude GST. Text to speech: ₹1.80 per 1,000 characters. Monthly commitments are credited to your wallet. Unused credit never expires. LLM: ₹15 input, ₹9.30 cached input and ₹55 output per 10,00,000 tokens. Pay as you go: 1,000 requests/minute, no monthly commitment. Business: 5,000 requests/minute with a ₹50,000 monthly commitment. Rate limits apply per workspace across all API keys. Monthly commitments are credited to your wallet. Unused credit never expires. All prices exclude GST. https://scalefrequency.com/pricing ## AI agents make and take your calls. They speak Hindi, English, Gujarati, Tamil and Telugu, and go live today. https://scalefrequency.com/ai-calling ## Your first API call takes five minutes. Fourier TTS speaks, Fourier STT listens and Mira LLM replies, all on one key. https://scalefrequency.com/developers ## Compared with other voice AI platforms (sources checked 8 and 17 October 2026) Our column comes from our price list. Every other value is quoted from that company's own public pages; a dash means the page does not say. ### Price per minute, as published - Frequency: ₹5/min on T5, covering speech, model, voice, phone calling and insights - Vapi: $0.05/min hosting; speech, model and transport billed separately - Retell AI: $0.07–$0.31/min for voice agents - Bland AI: $0.14/min on Start, covering model, speech and voice - Bolna: 6.00¢/min standard rate (₹5.52 shown), plus a $0.02/min platform fee - Gnani.ai: — - Skit.ai: — ### Hindi and other Indian languages - Frequency: Hindi, English, Gujarati, Tamil, Telugu and Hinglish - Vapi: Hindi, among 30+ languages; 40+ in Vapi Voices version 2 - Retell AI: Hindi, Kannada, Marathi, Tamil and Urdu; varies by provider - Bland AI: Hindi, also in Latin script; 40+ languages on the home page - Bolna: 18 languages, including Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu - Gnani.ai: Hindi, Tamil, Telugu, Kannada, Malayalam, Bengali, Gujarati, Marathi, Punjabi - Skit.ai: Hindi, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati, Bengali, Punjabi ### Call audio processed in India - Frequency: Yes: Mumbai, Noida and Delhi - Vapi: No India region named; US and EU infrastructure - Retell AI: No hosting region named - Bland AI: No India region named; SIP edges in the US, Canada, Asia-Pacific and Europe - Bolna: On enterprise plans; US by default - Gnani.ai: — - Skit.ai: — ### Phone calling included in the price - Frequency: Yes on T5 and T3; optional on T1 at ₹0.30/min - Vapi: No: calls billed by the provider - Retell AI: No: $0.15–$0.25/min for calls to India - Bland AI: Pages differ: “billed separately” (pricing), included (home) - Bolna: No: charged per minute by the carrier - Gnani.ai: — - Skit.ai: — ### Phone numbers - Frequency: Indian (+91) numbers: buy a dedicated number through us, or connect your own Plivo number - Vapi: One free US number for inbound calls; others imported from Twilio, Telnyx, DIDWW or a SIP trunk - Retell AI: US and Canada only ($2/month); other countries through custom telephony - Bland AI: Your Twilio account, any SIP trunk, or Bland’s built-in Twilio at pass-through cost - Bolna: — - Gnani.ai: — - Skit.ai: — ### Start without talking to sales - Frequency: Yes: look around with no sign-up, ₹200 credit to start - Vapi: Yes, $5 free credit - Retell AI: Yes, $10 free credit - Bland AI: Yes, Start plan, no card - Bolna: Yes, $5 free credit - Gnani.ai: — - Skit.ai: No: book a demo ### Time to go live, as stated - Frequency: Same day: go live in the console and call today - Vapi: “Build, test, and deploy in minutes” - Retell AI: “Go live in minutes” - Bland AI: Pages differ: “within a day” (pricing), “two to six weeks” (home) - Bolna: “Go from idea to live calls in minutes” - Gnani.ai: “<20mins to go-live” for APIs; agents “deployed in days” - Skit.ai: “Typical pilots go live in 4–6 weeks” Sources: https://scalefrequency.com/compare#sources ## Running in production ### Build the fleet We started by renting GPU capacity and building our fleet in Indian data centres. Then we took charge of how that compute runs, scales and serves every call. Rented compute. Operated by us. ### Own the core TTS is the most compute-intensive and expensive layer. We trained Fourier TTS from scratch and wrote its inference engine to deliver 3× the throughput of the industry standard. 24 streams per GPU, compared with the industry’s 8. ### Prove it live Before opening to the public, we spent three months running live with beta customers. Real calls, real feedback. Test, fix, test again. Three months with beta customers before public access. ### Stay bootstrapped Bootstrapping gives us room to focus on the core: better models, faster inference and lower costs. We build for customers, on our own terms. Customer revenue goes back into the core. ## Fourier TTS and languages Fourier TTS is live in production with customers. Hindi, English, Gujarati, Tamil and Telugu are live in speech recognition and text to speech, with the same four voices, and so is Hinglish. Bengali, Marathi and Kannada follow after evaluation. Benchmarks quoted on this site are Hindi. More information about the models is coming soon. For the available API, consult the current documentation. https://docs.miraiminds.co/v2/tts-quickstart ## All this infrastructure. One API call. The GPU fleet, the speech models, the inference engine and all the orchestration. We bring it together and keep it running. You build the conversation. ## A note from the founder I left my job in Silicon Valley to work on voice full time. For nine months we fought the same fight as everyone else — making agents as human and as accurate as we could. Then the market handed us a reality check: at ₹5 a minute, voice AI does not sell in India. So we went after the cost. The only honest way was to own the stack: our own GPUs, our own models, our own inference engine, our own SIP lines. That is how we made voice AI affordable at volume. What is behind this code is that stack, running live. Break it, push it, tell us what it gets wrong. — Sneh Mehta, founder ## Questions ### How does pricing work? Every company starts on T5: ₹5/minute, no monthly commitment. T3: ₹3/minute with a ₹50,000 monthly minimum. T1: ₹1/minute with a ₹1,00,000 monthly minimum, optional Mirai telco at ₹0.30/minute and optional insights at ₹0.20/minute. The same models and features, including single-prompt and node-based agents, are available on every tier. T5 and T3 include Mirai phone calling and insights. T1 adds ₹0.30/minute only for Mirai-managed phone calls. BYO telco, WhatsApp and browser calls have no Mirai telco charge on any tier. Provider charges you pay directly are separate. All prices exclude GST. ### How fast does it reply? About 850 ms on average, from the moment you finish speaking to the first reply audio: 590 ms for ASR, VAD and buffering, 150 ms for the LLM and 110 ms for TTS. It varies with speech, network and load, and the network distance to the caller is extra. ### Which languages? Hindi, English, Gujarati, Tamil and Telugu are live, in speech recognition and in the voice, and so is the Hinglish mix. Bengali, Marathi and Kannada follow. Each one ships when it clears evaluation, and the date is announced then. ### Where does the audio go? Voice processing runs in Indian data centres. You can review call recordings, transcripts and costs in the console. Website analytics and booking providers are covered separately in our privacy notice. ### Do I need my own telephony? T5 and T3 include Mirai phone calling and insights. T1 adds ₹0.30/minute only for Mirai-managed phone calls. BYO telco, WhatsApp and browser calls have no Mirai telco charge on any tier. Provider charges you pay directly are separate. ### Can you support a larger deployment? Yes. Talk to us about call volume, concurrency, workflows and the languages you need. We’ll work through capacity and commercial terms with you before you commit. ### How do I try it? Open the sandbox at sandbox.voice.miraiminds.co to build and test an agent. You can also start a conversation with the demo at the top of this page to hear Frequency in your browser. ### Is there an API? Yes — a v2 REST API with one auth header: create an agent, place a call, read the transcript, recording and cost. Fourier TTS also speaks OpenAI’s /v1/audio/speech. Everything is documented at docs.miraiminds.co. ## Developer guides ### Making Gemma 4 26B A4B fast enough for a voice conversation For a voice agent, optimize the time to useful output at your expected load. Our archived Gemma serving sweep recorded 138.1 ms p95 time to first content at 12 concurrent requests and 316.5 ms at 48. Those are client-side LLM measurements from short, warmed burst tests—not the caller’s full wait for speech. Here is the evidence and what it changes about deployment. https://scalefrequency.com/guides/gemma-4-26b-a4b-low-latency ### How FP4 can help voice inference—and what it cannot promise FP4 can reduce weight traffic and let supported hardware execute lower-precision operations. That can help LLM serving, but the latency gain depends on kernels, batch size, context and quality requirements. Our Gemma checkpoint uses an experts-only NVFP4 recipe; an FP4 label does not mean every tensor, or the KV cache, uses four bits. https://scalefrequency.com/guides/fp4-quantization-voice-inference ### Tune vLLM for voice latency, not just tokens per second A voice-serving configuration should maximize successful responses within a latency budget. Aggregate tokens per second alone can reward a configuration that makes each caller wait longer. Our short-response sweep increased estimated output from 803.2 to 2,103.0 units per second as concurrency rose from 12 to 48, while p95 first-content time increased from 138.1 to 316.5 ms. https://scalefrequency.com/guides/vllm-throughput-voice-agents ### Manage voice-agent context before buying more inference capacity Keep stable instructions at the beginning, store authoritative facts outside the transcript, and send only the history and tool data needed for the current turn. In our archived RTX 5090 tests, shared-prefix prompts recorded 225.7 ms p95 first-content time versus 501.2 ms for prompts with a unique leading nonce. That is an observed comparison, not a universal cache speedup. https://scalefrequency.com/guides/voice-agent-context-management ### Prompt-based or node-based voice agent: how to choose Use one prompt when the call does one job from start to finish. Use nodes when it has separate stages, branches, actions that need something first, or corrections you want to trace afterwards. Either way, keep permissions and side effects in your code, and let the model listen and talk inside each step. https://scalefrequency.com/guides/prompt-based-vs-node-based-voice-agents ### How to fix an overloaded voice-agent prompt An overloaded prompt asks the model to remember policy, track where the call is, decide on actions and keep a speaking style, all at once. Fix the contradictions, move the rules you can enforce into code, and give the model only what this turn needs. Then measure correctness and latency before you call it better. https://scalefrequency.com/guides/fix-overloaded-voice-agent-prompts ### Voice AI APIs for developers building in India There are four ways in, and you pick by the part of the call you'd rather not build. The voice-agent API places and answers whole phone calls for you. The Realtime API, Fourier TTS and transcription each take one piece, for teams that already run the rest. https://scalefrequency.com/guides/voice-ai-api-india ### Integrated voice API or build your own stack? Buy a hosted agent API if your product is the work around the call. Build your own pipeline if a particular component is your product and your team can run it. In between, keep your own orchestrator and hand each turn to a Realtime API. We sell all three, so you can start with one and move later. https://scalefrequency.com/guides/voice-ai-stack-build-vs-buy ### Choosing a Hindi and Hinglish text-to-speech API Fourier TTS, our own voice, is what our agents speak with, and you can call it on its own for ₹1.80 per 1,000 characters. Send Hindi in Devanagari with English words in Latin script, the way people type Hinglish, and pick one of four voices. Then judge it on your own names, amounts and mixed lines, played the way your users will hear them. https://scalefrequency.com/guides/hindi-text-to-speech-api ### Use Fourier TTS with an OpenAI-compatible speech client Already using OpenAI's speech API? Change three values and your code speaks Hindi, English, Gujarati, Tamil or Telugu with Fourier TTS. Just set response_format: we don't do MP3, and without it you get raw 48 kHz PCM. https://scalefrequency.com/guides/openai-compatible-tts ### Connect voice-agent webhooks to your application Your app hears about every call the moment it changes: started, finished, transcript ready. The receiver is one short Flask route that checks the signature, stores the event before it answers, and shrugs off retried duplicates. https://scalefrequency.com/guides/voice-agent-webhooks ### How to measure latency in a voice AI application The delay a caller feels runs from the end of their sentence to the first sound of the reply. The model's first token and the voice's first byte are slices of that wait, not the whole of it. Decide where each timer starts and stops, measure the full path on a real phone, and use the component numbers to see where the time went. https://scalefrequency.com/guides/voice-ai-latency ### Voice AI API pricing: minutes, characters and commitments Voice agents cost ₹1 a minute on the volume plan and ₹5 pay as you go, billed in 10-second blocks. Calls that never connect are free. Fourier TTS is billed by the character and transcription by the hour of audio, and everyone starts on pay as you go. https://scalefrequency.com/guides/voice-ai-api-pricing ### Make an AI phone call from Python with the Mirai API In ten minutes your Python script will place a real call in Hindi and get the transcript back. You create an agent once, place the call, and catch the result on a small Flask webhook. https://scalefrequency.com/guides/ai-phone-call-api-python ### Run bulk AI calls from a CSV with a Mirai campaign Upload a CSV this afternoon and an agent calls every row in Hindi, English, Gujarati, Tamil or Telugu, inside your calling hours. Missed calls get retried, spending stops at your budget, and each result comes back with your own columns attached. https://scalefrequency.com/guides/bulk-ai-calls-from-csv ### Hindi and Hinglish TTS in Pipecat with pipecat-mirai Swap in MiraiTTSService and your Pipecat bot speaks Hindi, English, Gujarati, Tamil, Telugu or Hinglish in one of four Fourier voices. It streams over one WebSocket per call, matches your pipeline's sample rate and stops when the caller cuts in. The same package can also replace your STT, LLM and TTS with one Realtime session. https://scalefrequency.com/guides/pipecat-hindi-tts ### Hindi speech-to-text API: transcribe calls with Mira Send us a recording of a Hindi or Hinglish call and get back who said what, and when. Hindi comes back in Devanagari and English in Latin script, as people mixed them, for ₹18 an hour of audio. https://scalefrequency.com/guides/hindi-speech-to-text-api ### Hindi text-to-speech APIs compared: prices and features, October 2026 Two Hindi text-to-speech APIs price in rupees: our Fourier TTS at ₹1.80 per 1,000 characters, and Sarvam's Bulbul from ₹2.00 on pay as you go, less on its monthly plans. ElevenLabs, Google, Azure, Amazon Polly, OpenAI and Cartesia price in US dollars. Here is every published price as of 16 October 2026, each with its source, and a way to test the voices on your own lines. https://scalefrequency.com/guides/hindi-tts-api-comparison ### Hindi speech-to-text APIs compared: prices and features, October 2026 Hindi speech to text is sold by the minute or by the hour. On 16 October 2026, we transcribe recorded Hindi and Hinglish calls at ₹18 an hour of audio, Sarvam lists ₹0.49 a minute, and six others price in US dollars. In our own report on RinggAI's public Hindi test, Fourier STT had the lowest error rate of the five systems scored, at 9.12%, with limits we explain below. Still, the only score that settles it is one on your own recordings. https://scalefrequency.com/guides/hindi-speech-to-text-api-comparison ### Realtime speech-to-speech API for Hindi voice agents Open one WebSocket, send the caller's audio, and get the agent's spoken reply back in Hindi, English, Gujarati, Tamil, Telugu or Hinglish. It speaks the OpenAI Realtime protocol, so your existing client works with a new key and URL. https://scalefrequency.com/guides/realtime-speech-to-speech-api ### Voice AI infrastructure in India: TTS, STT, LLM and agent APIs A voice AI product has four layers: speech to text, a language model, text to speech, and the agent platform that runs the call. Most companies sell one or two of them. Frequency and Sarvam sell all four, in rupees. Here's who sells what, and where each one says your audio is processed, as published on 16 October 2026. https://scalefrequency.com/guides/voice-ai-providers-india ## Use cases ### Confirm every COD order before it ships Every COD order gets a call before it leaves your warehouse. The agent reads back the order and the address in the customer's own language, and you get a yes, a no or the right address before you pack the box. Outbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: E-commerce, D2C brands. Cost per connected call at an assumed 80 seconds (80 s billed in 10-second blocks): T1 ₹1.34, T3 ₹4, T5 ₹6.67. All prices exclude GST. https://scalefrequency.com/use-cases/cod-order-confirmation ### Win back the carts that stopped at a question Most carts stop at one question. Will it fit? Is COD available here? The agent calls a few hours later, answers from your store's facts in the shopper's language, and offers to send the checkout link. Outbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: E-commerce, D2C brands. Cost per connected call at an assumed 110 seconds (110 s billed in 10-second blocks): T1 ₹1.84, T3 ₹5.50, T5 ₹9.17. All prices exclude GST. https://scalefrequency.com/use-cases/abandoned-cart-recovery ### Save failed deliveries before they become returns The courier says nobody was home. The customer says nobody came. The agent calls the same day, hears what happened, and settles it: a new day for the next attempt, a landmark the rider can find, or a cancellation so the parcel comes back. Outbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: Logistics, E-commerce, D2C brands. Cost per connected call at an assumed 100 seconds (100 s billed in 10-second blocks): T1 ₹1.67, T3 ₹5, T5 ₹8.34. All prices exclude GST. https://scalefrequency.com/use-cases/ndr-delivery-failure-calls ### Call every new lead while it's still warm Someone fills your form at 5:40 in the evening. By 8 they may be talking to two other builders. The agent calls while they're still thinking about you, asks the three or four questions your sales team needs, and books the visit. Outbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: Real estate, B2B services, Home services. Cost per connected call at an assumed 180 seconds (180 s billed in 10-second blocks): T1 ₹3, T3 ₹9, T5 ₹15. All prices exclude GST. https://scalefrequency.com/use-cases/lead-qualification ### Know who will pay on time, days before the due date Most customers know the EMI is due and say so in one sentence. A few have a reason the date won't work. The agent calls a few days early, confirms the amount and the date, and writes down when they'll pay, so your team only calls the people who need a person. Outbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: Lending and NBFCs, Subscriptions, B2B invoicing. Cost per connected call at an assumed 90 seconds (90 s billed in 10-second blocks): T1 ₹1.50, T3 ₹4.50, T5 ₹7.50. All prices exclude GST. https://scalefrequency.com/use-cases/payment-reminders ### Know by evening who's coming in tomorrow A reminder SMS gets read and ignored, and the chair sits empty. The agent calls each patient the day before, confirms the time, and moves the booking on the call if they can't make it. By evening, tomorrow's list is right. Outbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: Clinics and diagnostics, Salons and spas, Service centres. Cost per connected call at an assumed 70 seconds (70 s billed in 10-second blocks): T1 ₹1.17, T3 ₹3.50, T5 ₹5.84. All prices exclude GST. https://scalefrequency.com/use-cases/appointment-reminders ### Turn the monthly due list into booked service slots Your dealer system prints the due list every month, whether anyone calls it or not. The agent calls each owner, says what's due and until when, and books a slot at their usual centre. When they say no, you find out why. Outbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: Two-wheeler and car dealers, Appliances and water purifiers, Home services. Cost per connected call at an assumed 120 seconds (120 s billed in 10-second blocks): T1 ₹2, T3 ₹6, T5 ₹10. All prices exclude GST. https://scalefrequency.com/use-cases/service-due-reminders ### Find out why a customer gave you a 3 A 3 out of 5 on a form leaves you guessing what went wrong. The agent calls the day after a visit, asks for the score and then why, and keeps the answer in the customer's own words. If the reason is a problem still waiting to be fixed, your team hears about it. Outbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: Home services, E-commerce, Healthcare. Cost per connected call at an assumed 120 seconds (120 s billed in 10-second blocks): T1 ₹2, T3 ₹6, T5 ₹10. All prices exclude GST. https://scalefrequency.com/use-cases/customer-feedback-surveys ### Answer the phone when nobody's at the desk The phone rings while your receptionist is with a patient. The agent picks up, answers the questions you hear every day, and takes a proper message when the caller needs a person: their name, the reason and a good time to call back. Inbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: Clinics, Local services, Small businesses. Cost per connected call at an assumed 150 seconds (150 s billed in 10-second blocks): T1 ₹2.50, T3 ₹7.50, T5 ₹12.50. All prices exclude GST. https://scalefrequency.com/use-cases/inbound-receptionist ### Catch renewals before the policy lapses Most policies lapse because someone forgot. The agent calls a couple of weeks before expiry with the date and the amount, and offers the renewal link. When the customer asks why the premium went up, it books an advisor instead of guessing. Outbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: Insurance distribution, Subscriptions, Gyms and memberships. Cost per connected call at an assumed 150 seconds (150 s billed in 10-second blocks): T1 ₹2.50, T3 ₹7.50, T5 ₹12.50. All prices exclude GST. https://scalefrequency.com/use-cases/renewal-reminders ### Screen hundreds of applicants without a day on the phone Two hundred people applied for delivery jobs this week, and each one needs the same five-minute call. The agent calls them all, asks every candidate the same questions and tells them what happens next. Your recruiters get a sheet they can sort, and they make every decision. Outbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: Staffing agencies, Logistics and delivery, Retail and call centres. Cost per connected call at an assumed 210 seconds (210 s billed in 10-second blocks): T1 ₹3.50, T3 ₹10.50, T5 ₹17.50. All prices exclude GST. https://scalefrequency.com/use-cases/recruitment-screening ### Call back every enquiry before another institute does Results come out and a hundred enquiries land in one day. The agent calls each family back, answers the fee and batch questions you've given it, and books a slot with a counsellor. The counsellor starts that call knowing who picked up and what's still unanswered. Outbound and inbound calls in Hindi, English, Hinglish, Gujarati, Tamil, Telugu. Industries: Coaching institutes, Colleges and universities, Online courses. Cost per connected call at an assumed 240 seconds (240 s billed in 10-second blocks): T1 ₹4, T3 ₹12, T5 ₹20. All prices exclude GST. https://scalefrequency.com/use-cases/admissions-enquiry-follow-up ## Alternatives ### Looking for a Sarvam alternative? Sarvam is a strong choice if you need many Indian languages today. If your calls are in Hindi, Gujarati, Tamil, Telugu, English or a mix, we're built for exactly that, on our own models and GPUs in India. Our voice, on its own, is ₹1.80 per 1,000 characters. - Voice agents, published price per minute: Sarvam AI: Sarvam Voice Agents ₹3.50 pay as you go, ₹3.25 Pro, ₹3.15 Growth, ₹3.00 Business (coming soon), in 60-second pulses (https://www.sarvam.ai/products/voice-agents/pricing). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month - The agent's phone line: Sarvam AI: Extra: Sarvam telephony (Vobiz) ₹0.20 per 30-second pulse, numbers ₹59 per 30 days, or your own carrier at no Sarvam charge; one docs page says telephony is included (https://www.sarvam.ai/products/voice-agents/pricing). Frequency: Included on T5 and T3; ₹0.30/min on T1; your own Plivo number carries no Mirai telco charge - Text to speech, pay as you go, per 1,000 characters: Sarvam AI: Bulbul V4 Flash ₹2.00; Bulbul V3 ₹3.00 (https://www.sarvam.ai/api-pricing). Frequency: Fourier TTS ₹1.80 (GST extra) - Text to speech on a plan, per 1,000 characters: Sarvam AI: Bulbul V4 Flash ₹1.70 and Bulbul V3 ₹2.70 on Growth and Business; ₹1.85 and ₹2.85 on Pro (https://www.sarvam.ai/api-pricing). Frequency: ₹1.80 on Business (₹50,000/month), the same rate as pay as you go - Languages: Sarvam AI: Agents: 11 Indian languages and English; Bulbul V4 Flash covers Hindi, English, English–Hindi and seven more Indian languages; Saaras transcribes 23 (https://docs.sarvam.ai/api/getting-started/models/bulbul). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Speech to text: Sarvam AI: ₹0.49/min real-time or batch, pay as you go, on the pricing page; the docs pricing page lists ₹30 an hour (https://www.sarvam.ai/api-pricing). Frequency: Recorded calls ₹18 an hour of audio, speaker labels included; streaming recognition as a standalone API in private preview - API shape: Sarvam AI: Sarvam's own REST, HTTP streaming and WebSocket APIs, with Python and JavaScript SDKs (https://docs.sarvam.ai/api/getting-started/models/bulbul). Frequency: OpenAI's POST /v1/audio/speech protocol, a WebSocket endpoint and a Pipecat integration Checked 17 October 2026. https://scalefrequency.com/alternatives/sarvam ### Looking for an ElevenLabs alternative for India? ElevenLabs makes excellent voices in more than 90 languages, and if you need that range, it's the obvious pick. If your callers are in India and mix English into their own language, we're built for exactly that, and we bill in rupees. Our voice, on its own, costs ₹1.80 per 1,000 characters. - Published price per minute: ElevenLabs: Voice agents: $0.08 per additional call minute after plan minutes; LLM billed separately; burst calls $0.16/min (https://elevenlabs.io/pricing/agents). Frequency: T5 ₹5/min pay as you go, phone line, model and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - The language model: ElevenLabs: Billed separately, at the provider's published rate; the calculator's default, Gemini 2.5 Flash $0.0029/min (https://elevenlabs.io/pricing/agents). Frequency: The open Gemma 4 26B, served on our own GPUs in India, inside the per-minute rate on every tier - Agent plans: ElevenLabs: Free 15 minutes; Starter $6 for 75 minutes up to Business $990 for 12,375 minutes a month, with 4 to 40 concurrent calls (https://elevenlabs.io/pricing/agents). Frequency: No minute bundles; T3 and T1 commitments are credited to your wallet; 5 concurrent calls by default, raised on request - Telephony: ElevenLabs: No ElevenLabs telephony fee; your provider (Twilio or a SIP trunk) bills you directly (https://elevenlabs.io/pricing/agents). Frequency: Mirai phone calling included on T5 and T3, ₹0.30/min on T1; or connect your own Plivo number in India - Text to speech, per 1,000 characters: ElevenLabs: $0.04 for v4 Turbo, v3 Conversational and Flash/Turbo; $0.08 for v4, v3 and v2 Multilingual (https://elevenlabs.io/pricing/api). Frequency: Fourier TTS ₹1.80, on pay as you go and on the ₹50,000/month Business plan - Speech to text: ElevenLabs: Scribe v2 $0.22 an hour; Scribe v2 Realtime $0.39 an hour; 90+ languages (https://elevenlabs.io/pricing/api). Frequency: Recorded calls ₹18 an hour of audio, speaker labels included; streaming recognition as a standalone API in private preview - Languages: ElevenLabs: Eleven v4 lists 90+ languages, including Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Punjabi, Tamil and Telugu (https://elevenlabs.io/docs/overview/models). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Where it runs: ElevenLabs: Data stored in the US by default; EU, India and Singapore storage on Enterprise, where processing may still happen elsewhere (https://elevenlabs.io/docs/overview/administration/data-residency). Frequency: Every call runs on our own GPUs in India Checked 17 October 2026. https://scalefrequency.com/alternatives/elevenlabs ### Looking for a Vapi alternative for Indian calls? Vapi is a great fit if you like choosing every part of your agent yourself and calling numbers around the world. If your customers are in India, we sell the whole call as one price in rupees, with the phone line included on pay as you go. - Published price per minute: Vapi: $0.05/min Vapi hosting, plus speech-to-text, model, text-to-speech and transport at each provider's listed price (https://vapi.ai/pricing). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - Their calculator's default: Vapi: 1,000 minutes for $82–$129 a month: $0.05/min hosting, Deepgram $0.0095 - $0.0099/min, OpenAI $0.0077 - $0.0452/min and ElevenLabs $0.0146 - $0.0238/min; transport $0 on Vapi's own US numbers (https://vapi.ai/pricing). Frequency: 1,000 minutes on T5: 1,000 × ₹5 = ₹5,000, phone line included - What the price includes: Vapi: Hosting only. Models pass through at cost, or bill you directly on your own keys; transport is charged by the provider (https://docs.vapi.ai/customization/provider-keys). Frequency: The phone line, speech recognition, the model, the voice and call insights on T5 and T3; on T1 the line and insights are optional at ₹0.30/min and ₹0.20/min - Languages: Vapi: Vapi Voices version 1 names Hindi among 30+ languages; version 2 supports 40+ with automatic detection (https://docs.vapi.ai/providers/voice/vapi-voices). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Where it runs: Vapi: Orchestration on Vapi's US and EU infrastructure; the EU region needs the Premier package; no India region named (https://docs.vapi.ai/security-and-privacy/data-flow). Frequency: Every call runs on our own GPUs in India - Phone numbers: Vapi: One free US number for inbound calls; outbound and international calls use numbers imported from Twilio, Telnyx, DIDWW or a SIP trunk (https://docs.vapi.ai/phone-calling). Frequency: Indian (+91) numbers: buy a dedicated number through us, or connect your own Plivo number Checked 17 October 2026. https://scalefrequency.com/alternatives/vapi ### Looking for a Bolna alternative? Bolna builds for Indian callers, just like us. It speaks more Indian languages than we do today, and lets you bring your own models. We're for teams who'd rather not assemble the parts: one rupee price for the whole call, billed in 10-second blocks. - Published price per minute: Bolna: 6.00¢ (₹5.52) a minute on the pricing page, $0.06/min in the docs, for speech, model and voice on preferred models; the page works out its rupee figure at ₹92 to the dollar (https://www.bolna.ai/docs/pricing/preferred-models). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - What else a minute adds: Bolna: $0.02/min platform fee (plus provider charges), and telephony charged per minute by the carrier, billed by Bolna or by your own provider (https://www.bolna.ai/docs/frequently-asked-questions). Frequency: Nothing on T5 and T3: the phone line and insights are inside the rate; on T1 they are optional at ₹0.30/min and ₹0.20/min - Lower rates: Bolna: Bigger top-ups lower the rate, to 4.51¢ a minute at $3K; the Pilot plan is $500 for 12,000 minutes, billed in 30-second pulses, up to 25 concurrent calls (https://www.bolna.ai/pricing). Frequency: T3 ₹3/min with ₹50,000 a month; T1 ₹1/min with ₹1,00,000 a month; 10-second blocks on every tier - Languages: Bolna: 18 languages in the docs, Indian ones including Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu (https://www.bolna.ai/docs/customizations/supported-languages). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Where it runs: Bolna: US-hosted (AWS us-east-1) by default; India (Mumbai) processing on enterprise plans (https://www.bolna.ai/docs/enterprise/data-residency). Frequency: Every call runs on our own GPUs in India - Your own model keys: Bolna: Bring your own keys for LLM, TTS and ASR providers (https://www.bolna.ai/pricing). Frequency: Agents run on the Frequency stack on every tier Checked 17 October 2026. https://scalefrequency.com/alternatives/bolna ### Looking for a Retell AI alternative for India? Retell AI gives you a lot of choice: pick the model, pick the voice, transfer calls, work in dozens of languages. If you're calling Indian customers from Indian numbers, we make that simpler. One rupee price covers the agent, and on pay as you go it covers the phone line too. - Published price per minute: Retell AI: $0.07–$0.31/min for AI voice agents, pay as you go (https://www.retellai.com/pricing). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - Their calculator's default: Retell AI: $0.150/min: GPT 5.4 $0.080/minute, Retell Frequency $0.055/minute, Retell Platform Voices $0.015/minute and telephony $0.000 on your own carrier (https://www.retellai.com/pricing). Frequency: One rate per tier for speech, model and voice; T5 and T3 add the phone line and insights at no extra charge; on T1 they are optional at ₹0.30/min and ₹0.20/min - Calls to India: Retell AI: India (Twilio) $0.15/min and India (Telnyx) $0.25/min in the pricing page's call-rate list; no charge for SIP trunking or custom telephony (https://www.retellai.com/pricing). Frequency: Included on T5 and T3; ₹0.30/min on T1 for Mirai-managed calls; no Mirai telco charge on your own telco, WhatsApp or browser calls - Phone numbers: Retell AI: Retell-managed numbers are US and Canada only ($2/month); numbers in other countries connect through custom telephony (https://docs.retellai.com/deploy/purchase-number). Frequency: Indian (+91) numbers: buy a dedicated number through us, or connect your own Plivo number - Indian languages: Retell AI: Hindi, Kannada, Marathi, Tamil and Urdu in the voice table; speech recognition for all but Hindi only on its Azure and Soniox engines (https://docs.retellai.com/build/language-support). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Where it runs: Retell AI: No hosting region named on the pricing, billing or compliance pages (https://docs.retellai.com/general/compliance). Frequency: Every call runs on our own GPUs in India Checked 17 October 2026. https://scalefrequency.com/alternatives/retell-ai ### Looking for a Bland AI alternative? Bland AI keeps pricing simple: one rate a minute, no token charges. We like that too. The difference is where you're calling. For calls to India, in Hindi, Gujarati, Tamil, Telugu or English, our rate is in rupees, and the phone line is included on pay as you go. - Published price per minute: Bland AI: Start $0.14/min with no platform fee; Build $0.12/min with a $299/month platform fee (https://www.bland.ai/pricing). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - What the rate covers: Bland AI: LLM, STT and TTS, on Bland's own models. Pages differ on telephony: the pricing page says "Telephony is billed separately"; the home page and the docs say one rate covers telephony (https://www.bland.ai/pricing). Frequency: Speech, model, voice, Mirai phone calling and call insights on T5 and T3; on T1, phone calling (₹0.30/min) and insights (₹0.20/min) are optional - Time to go live, as stated: Bland AI: Pages differ: a first agent "within a day" on self-serve (pricing page); production agents in "two to six weeks" (home page) (https://www.bland.ai/pricing). Frequency: Same day: complete Go live in the console and call today - Languages: Bland AI: Hindi and Hindi (Latin script) in the send-call language list; the home page says 40+ languages (https://docs.bland.ai/api-v1/post/calls). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Your own telephony: Bland AI: Your Twilio account, SIP trunks from any provider, or Bland's built-in Twilio at pass-through cost (https://www.bland.ai/pricing). Frequency: Your own Plivo number in India, or a dedicated +91 number bought through us - Where it runs: Bland AI: SIP edges in the United States, Canada, Asia-Pacific and Europe; on-prem or VPC on Enterprise; no India region named (https://docs.bland.ai/enterprise-features/SIP-integration). Frequency: Every call runs on our own GPUs in India Checked 17 October 2026. https://scalefrequency.com/alternatives/bland-ai ### Looking for a Synthflow alternative with a price you can read? Synthflow is an enterprise voice-agent platform with its own telephony, warm transfers and a long compliance list. Its pricing page carries one figure: contracts start at $30,000 a year, and the per-minute rate is in the agreement. We publish ours. Pay as you go is ₹5 a minute with the phone line inside, and steady volume takes it to ₹1. - Published price: Synthflow: Enterprise contracts start at $30,000 annually. Final pricing is scoped around call volume, concurrency, telephony setup, integrations, security needs and launch support; no per-minute rate is on the page (https://synthflow.ai/pricing). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - How a minute is counted: Synthflow: Calls are measured per second and partial seconds are rounded up to the nearest second; the seconds are added up across calls and each whole minute is a billing event; unanswered calls are recorded as a flat 5 seconds of usage; failed and user-cancelled calls are not billed (https://docs.synthflow.ai/usage). Frequency: 10-second blocks on every plan; unanswered, busy and failed calls are not billed - Phone numbers: Synthflow: U.S., Canadian, or Australian numbers bought inside the platform at $1.50 USD per month; importing your own number from Twilio, Telnyx, Vonage, RingCentral, Zoom or a custom SIP carrier requires an Enterprise account (https://docs.synthflow.ai/phone-numbers). Frequency: Indian (+91) numbers: buy a dedicated number through us, or connect your own Plivo number - Languages: Synthflow: Hindi is the one Indian language on the list, on its own and inside the Multilingual mode that switches between English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch; Synthflow STT is recommended for non-English languages (https://docs.synthflow.ai/about-supported-languages). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Where it runs: Synthflow: API regions Global, United States and European Union (api.eu.synthflow.ai), with EU and US hosting on the home page; no India region is named (https://docs.synthflow.ai/api). Frequency: Every call runs on our own GPUs in India - Transfers and concurrency: Synthflow: Cold transfer over SIP REFER and warm transfer with a message or an AI summary; concurrency is scoped with Synthflow in the contract, with no number on the page (https://docs.synthflow.ai/call-transfers). Frequency: No transfer to a person yet; a new workspace starts with 5 calls at a time, raised when you ask - Compliance: Synthflow: ISO 27001:2022, SOC 2, GDPR, HIPAA and PCI DSS V4.0.1 listed on the security page (https://docs.synthflow.ai/security). Frequency: Every call runs on our own GPUs in India; the FAQ in our docs says where each part runs Checked 25 October 2026. https://scalefrequency.com/alternatives/synthflow ### Looking for a LiveKit alternative that comes with the agent? LiveKit gives developers the plumbing: an open-source agent framework, WebRTC and SIP, and hosted models you pay for by the minute. You write and run the agent. We are the other side of that choice: speech recognition, model and voice of our own, on our GPUs in India, sold as one rupee price per minute with the phone line inside on pay as you go. - Published price per minute: LiveKit: Agent sessions: Build $0/mo with 1,000 minutes included; Ship $50/mo with 5,000 minutes included then $0.01 per min; Scale $500/mo with 50,000 minutes included then $0.01 per min; the models and the phone line are billed on top (https://livekit.com/pricing). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - What else a minute adds: LiveKit: Models through LiveKit Inference at each model's price, for example Deepgram Nova-3 (Multilingual) at $0.0058 a minute and Cartesia Sonic 3 at $50.00 per million characters on Build and Ship; third-party SIP minutes at $0.004 per min on Ship and $0.003 per min on Scale; and your carrier's own charges for the call (https://livekit.com/pricing/inference). Frequency: Nothing on T5 and T3: the phone line and insights are inside the rate; on T1 they are optional at ₹0.30/min and ₹0.20/min - Their worked example: LiveKit: LiveKit's own example of a phone call on Build or Ship comes to $0.0479 a minute: the agent session, its US telephony, the model, speech recognition, the voice and observability; a carrier's line to India is not in it (https://livekit.com/pricing). Frequency: 1,000 minutes on T5: 1,000 × ₹5 = ₹5,000, phone line included - Concurrent agent sessions: LiveKit: 5 on Build, 20 on Ship, Up to 600 on Scale, custom on Enterprise (https://livekit.com/pricing). Frequency: A new workspace starts with 5 calls at a time, raised when you ask, with no code change - Languages: LiveKit: Whatever the model you pick supports; Hindi is in the Inference lists for Deepgram Nova-3, AssemblyAI and Cartesia Sonic 3, and Sarvam is a plugin on your own key, not in Inference (https://docs.livekit.io/agents/models/stt/). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Where it runs: LiveKit: A global edge network with region pinning on Scale; SOC 2 Type II; PCI DSS and ISO 27001 are in progress; no India region is named on the pages we read (https://livekit.com/security). Frequency: Every call runs on our own GPUs in India - Phone numbers: LiveKit: LiveKit Phone Numbers are available only in the US, and only inbound calling is supported; calls to India go through a SIP trunk from Twilio, Telnyx, Exotel, Plivo or another carrier (https://docs.livekit.io/telephony/start/phone-numbers/). Frequency: Indian (+91) numbers: buy a dedicated number through us, or connect your own Plivo number - Your own code: LiveKit: Open source under Apache 2.0; you write the agent in Python or Node.js and host it on LiveKit Cloud or your own servers (https://docs.livekit.io/agents/). Frequency: Prompt or node-based agents built in the console, started from your code with POST /v2/calls Checked 25 October 2026. https://scalefrequency.com/alternatives/livekit ### Looking for a Twilio ConversationRelay alternative for India? Twilio ConversationRelay turns a phone call into text for your app and speaks your replies back, with Google, Deepgram, Amazon and ElevenLabs doing the listening and the speaking. You bring the language model and host the WebSocket. We sell the whole call: our own speech recognition, model and voice, on our GPUs in India, priced in rupees with the line inside on pay as you go. - Published price per minute: Twilio ConversationRelay: Conversation Relay $0.07 / min on the US voice pricing page; the page does not say whether speech recognition and text to speech are inside it, and the call's own minutes are billed separately (https://www.twilio.com/en-us/voice/pricing/us). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - The call leg to India: Twilio ConversationRelay: Outbound local $0.0699 / min and outbound mobile $0.0496 / min on the India voice pricing page; BYOC trunking $0.0040 / min each way if you bring your own carrier (https://www.twilio.com/en-us/voice/pricing/in). Frequency: Nothing on T5 and T3: the phone line is inside the rate; on T1 it is optional at ₹0.30/min - SIP trunking to India: Twilio ConversationRelay: Elastic SIP Trunking to India starts at $0.0456 / min, with Unlimited Call Concurrency included (https://www.twilio.com/en-us/sip-trunking/pricing/in). Frequency: Buy a +91 number from us, or connect your own Plivo number with no phone charge from us - Who runs each part: Twilio ConversationRelay: Twilio handles the speech recognition and the voice: the TTS provider is Google, Amazon or ElevenLabs, and the STT provider is Google or Deepgram; the language model is yours, behind a WebSocket you host (https://www.twilio.com/docs/voice/twiml/connect/conversationrelay). Frequency: Fourier STT, Gemma 4 26B and Fourier TTS, all ours, on our own GPUs - Indian languages: Twilio ConversationRelay: The default voices table lists hi-IN, ta-IN, te-IN, kn-IN, ml-IN, mr-IN and en-IN, with an ElevenLabs voice for Hindi and Tamil; the language can be switched mid-call by a WebSocket message, and multi-language detection needs Deepgram with ElevenLabs (https://www.twilio.com/docs/voice/conversationrelay/voice-configuration). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Calling rules in India: Twilio ConversationRelay: Twilio's India guideline says "Outbound calls to India can only be made from international (non-Indian) numbers." (https://www.twilio.com/en-us/guidelines/in/voice). Frequency: Indian (+91) numbers: buy a dedicated number through us, or connect your own Plivo number - Where it runs: Twilio ConversationRelay: Twilio's default Region (US1) is located in the eastern United States; the regions page names no India region (https://www.twilio.com/docs/global-infrastructure/understanding-twilio-regions). Frequency: Every call runs on our own GPUs in India - Handing a call to a person: Twilio ConversationRelay: The end message with handoffData returns the call to Twilio, and your own TwiML dials the person (https://www.twilio.com/docs/voice/conversationrelay/websocket-messages). Frequency: Not yet; it is on our roadmap without a date Checked 25 October 2026. https://scalefrequency.com/alternatives/twilio-conversationrelay ### Looking for an OpenAI Realtime API alternative for Indian calls? The Realtime API is a model you build a phone agent around: one socket that listens and speaks, priced per audio token, with a SIP trunk and a server of your own. We sell the finished call. Hindi, Gujarati, Tamil, Telugu and English agents on our own stack in India, one price a minute in rupees, and the phone line inside it on pay as you go. - Published price per minute: OpenAI Realtime API: GPT-Live voice sessions are $0.05 per minute, billed per second, with the backend model and tool usage billed separately; the gpt-realtime models publish no per-minute price, only token rates (https://developers.openai.com/api/docs/models/gpt-live-1). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - Audio tokens: OpenAI Realtime API: gpt-realtime-2.1 audio: $32.00 input, $64.00 output and $0.40 cached input per 1M tokens; gpt-realtime-2.1-mini audio: $10.00 input and $20.00 output; a caller's audio is 1 token per 100 ms of audio and the assistant's is 1 token per 50ms of audio (https://developers.openai.com/api/docs/pricing). Frequency: No tokens to count: the whole call is one rate a minute, in 10-second blocks - What the price includes: OpenAI Realtime API: The voice model's tokens, or the GPT-Live session; the backend model, tools, the SIP trunk and the carrier's minutes are separate, and every turn resends the whole conversation, so turns later in the session will be more expensive (https://developers.openai.com/api/docs/guides/realtime-costs). Frequency: The phone line, speech recognition, the model, the voice and call insights on T5 and T3; on T1 the line and insights are optional at ₹0.30/min and ₹0.20/min - Languages: OpenAI Realtime API: No supported-language list is published for the Realtime API or GPT-Live; the prompting guide names Hindi only as an accent of English; ten voices: alloy, ash, ballad, coral, echo, sage, shimmer, verse, marin and cedar (https://developers.openai.com/api/docs/guides/realtime-conversations). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Where it runs: OpenAI Realtime API: India (in.api.openai.com) is listed for regional storage, not regional processing, and the Realtime API is not among the endpoints listed for it; OpenAI says support for regional storage does not imply support for regional processing (https://developers.openai.com/api/docs/guides/your-data). Frequency: Every call runs on our own GPUs in India - Phone numbers: OpenAI Realtime API: None of its own: calls arrive over SIP from a trunking provider such as Twilio, and outbound SIP calling must be enabled for your organisation; ringing is limited to 3 minutes, a connected call to 2 hours, and a Realtime session to 60 minutes (https://developers.openai.com/api/docs/guides/realtime-sip). Frequency: Indian (+91) numbers: buy a dedicated number through us, or connect your own Plivo number - Transfer to a person: OpenAI Realtime API: Yes, with a SIP REFER to a target of your choosing (https://developers.openai.com/api/docs/guides/realtime-sip). Frequency: Not yet; the agent can promise a callback, and your system picks it up from the call's outcome Checked 25 October 2026. https://scalefrequency.com/alternatives/openai-realtime ### Looking for a Dialogflow CX alternative for Indian calls? Dialogflow CX is Google's agent builder on Google's own speech and models, billed per second of audio, with a free trial and a Mumbai region. Its phone gateway stops at US numbers. We sell the whole call for callers in India: Hindi, Gujarati, Tamil, Telugu and English on our own GPUs, one price a minute in rupees, and the phone line inside it on pay as you go. - Published price per minute: Google Dialogflow CX: Voice agents are billed per second of audio: $0.001 / 1 second on Flows and $0.002 / 1 second on Playbooks; the page prices seconds of audio and does not itemise speech recognition, the voice or the model (https://cloud.google.com/products/conversational-agents/pricing). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - How a turn is billed: Google Dialogflow CX: Per turn: the seconds the agent spends listening, including listening for an interruption while it speaks, plus the full length of its spoken reply, rounded up to the nearest second, even when the caller cuts the reply off (https://cloud.google.com/products/conversational-agents/pricing). Frequency: The whole call in 10-second blocks; unanswered, busy and failed calls are not billed - Lower rates: Google Dialogflow CX: No volume rate is printed; the page asks you to request a custom quote from sales (https://cloud.google.com/products/conversational-agents/pricing). Frequency: T3 ₹3/min with ₹50,000 a month; T1 ₹1/min with ₹1,00,000 a month; 10-second blocks on every tier - Languages: Google Dialogflow CX: Hindi is on the Dialogflow CX language list with every feature ticked, and Bengali, Marathi, Gujarati, Kannada, Malayalam, Tamil, Telugu and Punjabi are on it too; Chirp 3 HD voices list Hindi (India) and nine more Indian locales (https://docs.cloud.google.com/dialogflow/cx/docs/reference/language). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Where it runs: Google Dialogflow CX: Agents can be created in Mumbai (asia-south1) and keep data at rest there, but the AI/ML Data Location commitment is only supported in the locations within the US and EU; once an agent is created, its location cannot change (https://docs.cloud.google.com/dialogflow/cx/docs/concept/region). Frequency: Every call runs on our own GPUs in India - Phone numbers: Google Dialogflow CX: Phone Gateway: support is limited to US phone numbers, and only for agents in the global region; elsewhere you bring a SIP trunk from your own session border controller, or use the AudioCodes, Avaya, Twilio or Voximplant integrations (https://docs.cloud.google.com/dialogflow/cx/docs/concept/integration/phone-gateway). Frequency: Indian (+91) numbers: buy a dedicated number through us, or connect your own Plivo number - Try it free: Google Dialogflow CX: $600 of trial credit for Flows and $1000 for Playbooks, expiring after 12 months (https://cloud.google.com/products/conversational-agents/pricing). Frequency: No sign-up to look around the sandbox; ₹200 of credit once you confirm a work email Checked 25 October 2026. https://scalefrequency.com/alternatives/dialogflow-cx ### Looking for an Amazon Connect alternative for Indian calls? Amazon Connect is a full contact centre: Connect runs the call, Amazon Lex understands it and Amazon Polly speaks, each billed on its own page. India is not in its default outbound lists, and Connect has no Mumbai endpoint. We sell one thing, the phone call in India: Hindi, Gujarati, Tamil, Telugu and English agents on our own GPUs, one price a minute in rupees, line included on pay as you go. - Published price per minute: Amazon Connect: Voice: $0.038 per voice min (standard telephony rates apply), billed per second with a minimum of 10 seconds; outbound campaign calls are $0.045 per minute instead (https://aws.amazon.com/products/connect/customer/pricing/). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - What else a minute adds: Amazon Connect: Telephony at Amazon's per-country rates, billed per second with a 60-second minimum; India sits in its Shared Trunking list, where you contract a telco partner directly and the carrier bills the number usage (https://aws.amazon.com/products/connect/customer/pricing/appendix/). Frequency: Nothing on T5 and T3: the phone line and insights are inside the rate; on T1 they are optional at ₹0.30/min and ₹0.20/min - Lower rates: Amazon Connect: No volume rate is printed for Connect voice; the page offers a pricing quote, and says no minimums, no long-term contracts (https://aws.amazon.com/products/connect/customer/pricing/). Frequency: T3 ₹3/min with ₹50,000 a month; T1 ₹1/min with ₹1,00,000 a month; 10-second blocks on every tier - Speech and the voice on their own: Amazon Connect: Amazon Lex is $0.004 per speech request, or $0.0065 per 15-second interval of streaming, silence included; Amazon Polly is $4.00 per 1 million characters for Standard voices, $16.00 for Neural, $30 for Generative and $100.00 for Long-Form (https://aws.amazon.com/lex/pricing/). Frequency: Inside every agent; on their own, Fourier TTS is ₹1.80 per 1,000 characters and Fourier Transcribe ₹18 an hour of recorded audio - Languages: Amazon Connect: Hindi (hi_IN) in Amazon Lex and in Connect's AI agents and agentic voice, with Bengali, Kannada, Marathi, Tamil and Telugu for agentic voice; Amazon Polly has four Hindi voices: Aditi, Kajal, Arjun and Kiara (https://docs.aws.amazon.com/connect/latest/adminguide/supported-languages.html). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Where it runs: Amazon Connect: Amazon Connect's region list and Amazon Lex's endpoint list have no Mumbai (ap-south-1) entry; Amazon Polly does, at polly.ap-south-1.amazonaws.com (https://docs.aws.amazon.com/connect/latest/adminguide/regions.html). Frequency: Every call runs on our own GPUs in India - Phone numbers: Amazon Connect: Numbers in more than 100 countries at a per-day rate, but outbound calling is a default deny model and India is in no Region's default allowlist or in the outbound campaign country list; India is served through Shared Trunking with a telco partner (https://docs.aws.amazon.com/connect/latest/adminguide/country-code-allow-list.html). Frequency: Indian (+91) numbers: buy a dedicated number through us, or connect your own Plivo number - Concurrent calls: Amazon Connect: Concurrent active calls per instance defaults to 10, raised by a quota request, and concurrent campaign calls to 0 (https://docs.aws.amazon.com/connect/latest/adminguide/amazon-connect-service-limits.html). Frequency: 5 calls at a time on a new workspace, raised when you ask, with no code change Checked 25 October 2026. https://scalefrequency.com/alternatives/amazon-connect ### Looking for a Gnani.ai alternative with a price on the page? Gnani builds its own Indian-language speech, voices and models, and sells them to banks and insurers through a sales conversation, with on-premise and air-gapped options. Its website prints no price. We publish ours in rupees, open the sandbox without a sign-up, and put you live the same day. - Price and plans: Gnani.ai: Not published. gnani.ai/pricing redirects to a Book a Demo form, and the speech API, enterprise and docs pages print no price, free tier or billing increment (https://www.gnani.ai/pricing). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - Languages: Gnani.ai: Prisma v2.5 speech recognition supports 10 Indian languages in the docs: Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Punjabi, Tamil and Telugu; the marketing page adds Odia and code-switched combinations like Hinglish, Tanglish, Benglish; Timbre voices speak the same ten (https://docs.gnani.ai/api/STT/speech-to-text). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Hindi voices: Gnani.ai: Timbre v2.5 lists 13 Hindi voices and one Hinglish voice; Timbre v3.0 has 13 voices that each speak all 10 languages (https://docs.gnani.ai/api/TTS/available-voices). Frequency: Four voices, neha, ashu, shruti and sameer, each speaking Hindi, English, Gujarati, Tamil and Telugu - Where it runs: Gnani.ai: Cloud, private cloud, on-premise, and air-gapped deployment, citing RBI and IRDAI residency rules; the agent platform's region list in the docs is Asia, Americas and Europe, with no city named (https://www.gnani.ai/). Frequency: Every call runs on our own GPUs in India - Telephony: Gnani.ai: Import a Twilio number into an agent, test calls to whitelisted numbers, a Webex Contact Center integration; the enterprise page lists Voice, WhatsApp, Chat, and SMS and transfer to a human agent with full conversation context; no +91 number rental or Indian carrier is named in the docs (https://www.gnani.ai/enterprise). Frequency: Indian (+91) numbers: buy a dedicated number through us, or connect your own Plivo number; WhatsApp and browser calling with no telco charge; no transfer to a person yet - Compliance: Gnani.ai: SOC 2 certified, ISO 27001 certified, and compliant with GDPR, HIPAA, and PCI-DSS, as its home page says (https://www.gnani.ai/). Frequency: We publish where every call runs and what we keep; ask us for what your review needs - Try it: Gnani.ai: Get API Key and Start Building lead to a sign-in dashboard; the agent platform is sold through a demo call (https://www.gnani.ai/book-demo). Frequency: No sign-up to look around the sandbox; ₹200 of credit once you confirm a work email; live the same day - Streaming and formats: Gnani.ai: Public docs: REST speech to text up to 60 seconds a request, WebSocket and SSE streaming, batch jobs, voice cloning, a Python SDK, LiveKit and Pipecat plugins; 8 kHz μ-law and A-law output for phone lines (https://docs.gnani.ai/api/TTS/tts-inference). Frequency: POST /v2/calls and signed webhooks for agents; OpenAI's speech API for Fourier TTS; POST /v2/transcriptions for recordings; 8 kHz μ-law or A-law Checked 25 October 2026. https://scalefrequency.com/alternatives/gnani ### Looking for an OpenAI TTS alternative with Indian voices? OpenAI's speech API is simple to use, and its own guide says the voices are optimised for English. If your users speak Hindi, Gujarati, Tamil or Telugu, Fourier TTS takes the same request, so your code stays. Change the base URL and the model name, and you get voices that slip into English naturally, at ₹1.80 per 1,000 characters. - Published price: OpenAI TTS: gpt-4o-mini-tts: $0.60 per 1M text input tokens and $12.00 per 1M audio output tokens. tts-1: $15.00 per 1M characters; tts-1-hd: $30.00 per 1M characters (https://developers.openai.com/api/docs/pricing). Frequency: ₹1.80 per 1,000 characters, on pay as you go and on Business (₹50,000/month commitment), before GST - Model status: OpenAI TTS: tts-1, tts-1-hd and gpt-4o-mini-tts are deprecated and will be removed from the API on January 6, 2027; OpenAI points to gpt-realtime-2.1-mini (https://developers.openai.com/api/docs/deprecations). Frequency: Fourier TTS is current, and it's the voice our own phone agents speak with - Indian languages: OpenAI TTS: The TTS model "generally follows the Whisper model" for languages, a list that includes Hindi, Kannada, Marathi and Tamil; the guide says voices are "optimized for English" (https://developers.openai.com/api/docs/guides/text-to-speech). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish; voices that switch into English mid-sentence - Phone-line audio: OpenAI TTS: MP3 (default), Opus, AAC, FLAC, WAV, and PCM at 24 kHz; no μ-law, A-law or 8 kHz (https://developers.openai.com/api/docs/guides/text-to-speech). Frequency: PCM or WAV, 48 kHz by default, plus raw 8 kHz μ-law and A-law for phone lines. MP3 is not offered - Streaming: OpenAI TTS: Chunked transfer encoding, so audio can play before the whole file is generated (https://developers.openai.com/api/docs/guides/text-to-speech). Frequency: Raw PCM streams over HTTP as it renders; a WebSocket endpoint keeps one connection open for a whole call - How to switch: OpenAI TTS: POST /v1/audio/speech with model, voice, input and response_format (https://developers.openai.com/api/docs/guides/text-to-speech). Frequency: The same POST /v1/audio/speech shape. Set the base URL to our /v1 and the model to mira-tts, and always send response_format - Where it runs: OpenAI TTS: India region (in.api.openai.com) lists storage, not processing, for /v1/audio/speech; data residency is for approved customers via sales (https://developers.openai.com/api/docs/guides/your-data). Frequency: Every call and every request runs on our own GPUs in India Checked 17 October 2026. https://scalefrequency.com/alternatives/openai-tts ### Looking for a Google Cloud TTS alternative for Hinglish? Google gives you a big shelf of voices in many Indian languages and a generous free tier, and if you already run on Google Cloud, that's convenient. Our voices are built for the way people speak on calls, slipping into English and back. They're made on our GPUs in India and cost ₹1.80 per 1,000 characters, billed in rupees. - Published price: Google Cloud Text-to-Speech: Chirp 3: HD voices US$30 per 1M characters; Neural2 US$16; WaveNet and Standard US$4; Gemini-TTS priced per token (https://cloud.google.com/text-to-speech/pricing). Frequency: ₹1.80 per 1,000 characters, on pay as you go and on Business (₹50,000/month commitment), before GST - Try it free: Google Cloud Text-to-Speech: First 1M Chirp 3: HD and Neural2 characters and first 4M WaveNet and Standard characters each month; billing must be enabled (https://cloud.google.com/text-to-speech/pricing). Frequency: No sign-up to look around the sandbox; ₹200 of credit once you confirm a work email; clips you play in the console are not charged - Voices: Google Cloud Text-to-Speech: hi-IN: 30 Chirp 3: HD voices, 4 Neural2, 6 WaveNet and 6 Standard; 30 Chirp 3: HD voices each for Gujarati, Tamil, Telugu, Bengali, Marathi, Kannada and Malayalam (https://cloud.google.com/text-to-speech/docs/list-voices-and-types). Frequency: Four voices, ashu, neha, shruti and sameer, in Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish - Streaming: Google Cloud Text-to-Speech: Two-way streaming synthesis, in Preview, for Chirp 3: HD only; streaming output is ALAW, MULAW, OGG_OPUS or PCM, and SSML is not supported when streaming (https://cloud.google.com/text-to-speech/docs/create-audio-text-streaming). Frequency: Raw PCM streams over HTTP as it renders; a WebSocket endpoint keeps one connection open for a whole call - How to switch: Google Cloud Text-to-Speech: Google's Cloud Text-to-Speech API and client libraries (https://cloud.google.com/text-to-speech/docs/chirp3-hd). Frequency: OpenAI's POST /v1/audio/speech: any OpenAI SDK works with our base URL and model mira-tts - Where it runs: Google Cloud Text-to-Speech: Global, US and EU endpoints. In Asia Pacific, Chirp 3: HD is marked for Tokyo and Singapore; the table lists Mumbai with no model marked (https://cloud.google.com/text-to-speech/docs/endpoints). Frequency: Every call and every request runs on our own GPUs in India Checked 17 October 2026. https://scalefrequency.com/alternatives/google-cloud-tts ### Looking for an Azure Speech alternative for Indian calls? Azure Speech is a broad, safe choice, with many Indian languages and a region in Central India. We're narrower and closer to the phone call. Our voices switch into English inside a sentence, our transcripts keep each word in its own script, and both bill one rupee wallet: ₹1.80 per 1,000 characters to speak, ₹18 an hour to transcribe. - Text to speech price: Azure Speech: Neural and Neural HD Flash: $15 per 1M characters; Neural HD and MAI-Voice: $22 per 1M characters (https://azure.microsoft.com/en-us/pricing/details/speech/). Frequency: ₹1.80 per 1,000 characters, on pay as you go and on Business (₹50,000/month commitment), before GST - Speech to text price: Azure Speech: Real-time $1 per hour; fast transcription $0.36; batch $0.18; speaker diarization on real-time $0.3 per hour per feature (https://azure.microsoft.com/en-us/pricing/details/speech/). Frequency: Recorded calls: ₹18 per hour of audio (₹0.30 a minute), speaker labels included, charged once when the job completes. Streaming recognition is in private preview as a standalone API - Indian languages: Azure Speech: hi-IN neural voices (Swara, Madhur, Aarav, Ananya, Kavya and others) and MAI-Voice voices in preview; two neural voices each for Gujarati, Tamil, Telugu, Bengali, Marathi, Kannada, Malayalam and Punjabi (https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - How to switch: Azure Speech: Azure's Speech SDK and REST APIs, with an endpoint per region (https://learn.microsoft.com/en-us/azure/ai-services/speech-service/regions). Frequency: TTS on OpenAI's POST /v1/audio/speech; transcription on POST /v2/transcriptions; one sk_live_ key for both - Where it runs: Azure Speech: Central India supports neural TTS and real-time, fast and batch transcription; South India is not supported for standalone speech (https://learn.microsoft.com/en-us/azure/ai-services/speech-service/regions). Frequency: Every call and every request runs on our own GPUs in India - Try it free: Azure Speech: Free F0 tier: 0.5 million neural characters and 5 audio hours of real-time transcription a month (https://azure.microsoft.com/en-us/pricing/details/speech/). Frequency: No sign-up to look around the sandbox; ₹200 of credit once you confirm a work email; clips you play in the console are not charged Checked 17 October 2026. https://scalefrequency.com/alternatives/azure-speech ### Looking for an Amazon Polly alternative for Hindi? If you use Polly for Hindi, your voice is Aditi or Kajal. They're solid bilingual voices, and Polly serves Kajal from Mumbai. Fourier TTS gives you four newer voices, made for phone calls where English words come up all the time, in Hindi, Gujarati, Tamil, Telugu and English, at ₹1.80 per 1,000 characters from a rupee wallet. - Published price: Amazon Polly: Standard $4.00, Neural $16.00, Generative $30 and Long-Form $100.00 per 1 million characters (https://aws.amazon.com/polly/pricing/). Frequency: ₹1.80 per 1,000 characters, on pay as you go and on Business (₹50,000/month commitment), before GST - Hindi voices: Amazon Polly: hi-IN: Aditi (Standard) and Kajal (Neural); no Generative or Long-Form Hindi voice listed; only Hindi and Indian English among Indian languages (https://docs.aws.amazon.com/polly/latest/dg/available-voices.html). Frequency: Four voices, ashu, neha, shruti and sameer, in Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish - Hinglish: Amazon Polly: Aditi and Kajal are bilingual Hindi and Indian English voices that can switch within one sentence, in Devanagari or Romanised Hindi (https://docs.aws.amazon.com/polly/latest/dg/bilingual-voices.html). Frequency: English inside the sentence, in one voice; Hindi in Devanagari with English words in Latin script - Streaming and formats: Amazon Polly: SynthesizeSpeech returns an audio stream: mp3, ogg_opus, ogg_vorbis, pcm (8 or 16 kHz), mulaw or alaw; two-way streaming is generative voices only, and Hindi has none (https://docs.aws.amazon.com/polly/latest/APIReference/API_SynthesizeSpeech.html). Frequency: Raw PCM streams as it renders, 48 kHz by default; 8 kHz μ-law and A-law for phone lines; WAV files; a WebSocket for whole calls - Where it runs: Amazon Polly: Neural voices supported in Asia Pacific (Mumbai), ap-south-1, among other regions (https://docs.aws.amazon.com/polly/latest/dg/neural-voices.html). Frequency: Every call and every request runs on our own GPUs in India - Try it free: Amazon Polly: 5M Standard characters a month; 1M Neural characters a month for 12 months; new customers get up to $200 in AWS Free Tier credits (https://aws.amazon.com/polly/pricing/). Frequency: No sign-up to look around the sandbox; ₹200 of credit once you confirm a work email; clips you play in the console are not charged Checked 17 October 2026. https://scalefrequency.com/alternatives/aws-polly ### Looking for a Deepgram alternative for Indian calls? Deepgram is a speech-to-text company first, with a public streaming API, 45+ languages and $200 of free credit. We're narrower: phone calls in India. We turn a recording into a transcript that knows who's speaking and keeps Hindi in Devanagari and English in Latin script, for ₹18 an hour. And our voice speaks Hindi, which Deepgram's Aura voices don't list. - Pre-recorded price: Deepgram: Nova-3 Monolingual $0.0043/min; Nova-3 Multilingual $0.0052/min; Whisper Large $0.0048/min; speaker diarization included on pre-recorded audio (https://deepgram.com/pricing). Frequency: ₹18 per hour of audio (₹0.30 a minute), charged to the millisecond when the job completes; speaker labels included - Streaming price: Deepgram: Nova-3 Monolingual $0.0048/min (limited-time; regular $0.0077); Nova-3 Multilingual $0.0058/min (regular $0.0092); Flux Multilingual $0.0078/min; diarization $0.0020/min (https://deepgram.com/pricing). Frequency: Streaming recognition runs inside every voice agent; as a standalone API it is in private preview, on request - Indian languages: Deepgram: Nova-3 lists Hindi, Bengali, Gujarati, Kannada, Marathi, Punjabi, Tamil and Telugu; its multilingual mode covers Hindi with English and eight other languages (https://developers.deepgram.com/docs/models-languages-overview). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish; Hindi in Devanagari, English in Latin script - An Indian voice: Deepgram: Aura's language list: English, Spanish, German, French, Dutch, Italian and Japanese; Flux TTS voices speak English (https://developers.deepgram.com/docs/tts-models). Frequency: Fourier TTS: Hindi, Gujarati, English, Tamil and Telugu, ₹1.80 per 1,000 characters - Where it runs: Deepgram: An EU endpoint, and an India endpoint (api.in.deepgram.com) that Deepgram says is hosted in India; Whisper is not on it (https://developers.deepgram.com/reference/regional-endpoints). Frequency: Every call and every request runs on our own GPUs in India - Try it free: Deepgram: $200 free credit, then pay as you go; no credit card required (https://deepgram.com/pricing). Frequency: No sign-up to look around the sandbox; ₹200 of credit once you confirm a work email Checked 17 October 2026. https://scalefrequency.com/alternatives/deepgram ### Looking for a Cartesia alternative, billed in rupees? Cartesia's Sonic speaks 44 languages, clones voices and streams a line as your model writes it. It's sold on monthly credit plans, in dollars. Fourier TTS is simpler: voices for Indian calls that slip into English, ₹1.80 per 1,000 characters from a rupee wallet, and no plan to pick. Even pay as you go lets five requests run at once. - Published price: Cartesia: Free: 20K credits/month. Pro $5/month: 100K. Startup $49: 1.25M. Scale $299: 8M. One credit is one character for Sonic; overage $65 (Pro), $45 (Startup) and $38 (Scale) per 1M credits (https://cartesia.ai/pricing). Frequency: ₹1.80 per 1,000 characters, on pay as you go and on Business (₹50,000/month commitment), before GST - Hindi and Hinglish: Cartesia: Sonic 3.6 lists Hindi among 44 languages, with "expanded support for Hindi transcripts written in Latin script"; Cartesia documents Hindi–English code-switching in one generation (https://docs.cartesia.ai/build-with-cartesia/capability-guides/advanced-capabilities). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish; Hindi in Devanagari with English words in Latin script - Phone-line audio: Cartesia: Raw pcm_mulaw or pcm_alaw at 8 kHz; PCM up to 48 kHz; WAV and MP3 on the bytes endpoint only (https://docs.cartesia.ai/build-with-cartesia/capability-guides/tts-output-audio-format). Frequency: Raw 8 kHz μ-law or A-law; PCM or WAV, 48 kHz by default - Concurrent requests: Cartesia: 2 on Free, 3 on Pro, 5 on Startup, 15 on Scale; custom on Enterprise (https://cartesia.ai/pricing). Frequency: 5 on pay as you go, 25 on Business; more on Enterprise - How to switch: Cartesia: Cartesia's own API: POST /tts/bytes, POST /tts/sse and a TTS WebSocket (https://docs.cartesia.ai/use-the-api/compare-tts-endpoints). Frequency: OpenAI's POST /v1/audio/speech with model mira-tts, and a WebSocket for whole calls - Where it runs: Cartesia: The pricing page names no region; Enterprise customers can get dedicated deployments in regions including India (https://docs.cartesia.ai/enterprise/regional-endpoints). Frequency: Every call and every request runs on our own GPUs in India Checked 17 October 2026. https://scalefrequency.com/alternatives/cartesia ### Looking for a Smallest.ai alternative? Smallest sells its own speech, voice and agent runtime, in dollars, with a Mumbai region and more Indian languages than we have today. We sell the whole phone call as one price in rupees, with the line included on pay as you go, and our voice on its own at ₹1.80 per 1,000 characters. - Voice agents, published price per minute: Smallest.ai: Depending on the models $0.09 - $0.21/minute on pay as you go, for speech, model and voice on a cascade; the phone number or your carrier's line is on top (https://smallest.ai/pricing). Frequency: T5 ₹5/min pay as you go, phone line and insights included; T3 ₹3/min with ₹50,000/month; T1 ₹1/min with ₹1,00,000/month (GST extra) - What else a minute adds: Smallest.ai: Smallest hosting at $0.01/min, Flat Rate; its layer rates are ~$0.009/minute for speech to text and ~$0.09/minute for its voice, with the language model at cost (https://smallest.ai/pricing). Frequency: Nothing on T5 and T3: the platform, speech recognition, model, voice, phone line and insights are inside the rate; on T1 the line and insights are optional at ₹0.30/min and ₹0.20/min - Lower rates: Smallest.ai: Enterprise is as low as $0.05/minute, with custom hosting, dedicated infrastructure and a 99.99% SLA; pay as you go has no volume rate (https://smallest.ai/pricing). Frequency: T3 ₹3/min with ₹50,000 a month; T1 ₹1/min with ₹1,00,000 a month; 10-second blocks on every tier - Text to speech, pay as you go: Smallest.ai: Lightning V3.1 at ~$0.175/10K characters and Lightning V3.1 Pro at ~$0.195/10K characters, from a prepaid balance; Enterprise is custom (https://smallest.ai/pricing/models). Frequency: ₹1.80 per 1,000 characters, on pay as you go and on Business (₹50,000/month commitment), before GST - Speech to text: Smallest.ai: Pulse pre-recorded at ~$0.003/minute and Pulse real-time at ~$0.004/minute on the pricing page; the docs compare page lists Pulse at $0.005 per min batch and $0.008 per min streaming (https://smallest.ai/pricing/models). Frequency: ₹18 per hour of audio (₹0.30 a minute), speaker labels included, charged once when the job completes; streaming recognition as a standalone API in private preview - Languages: Smallest.ai: Lightning V3.1 accepts 10 Indic: Hindi, Marathi, Gujarati, Punjabi, Bengali, Odia, Tamil, Telugu, Kannada, Malayalam, with 217 voices total across 12 languages and Hindi (115) the biggest; Indian Pro voices speak English with Hindi code-switching (https://docs.smallest.ai/models/text-to-speech/voices-languages). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Where it runs: Smallest.ai: Two regions, India (Mumbai, api.india.smallest.ai) and USA (Oregon), with requests routed to the nearer one unless you pin a hostname; Tamil, Telugu, Kannada and Malayalam streaming recognition are Beta and India region only (https://docs.smallest.ai/api-reference/concurrency-and-limits). Frequency: Every call and every request runs on our own GPUs in India - Phone numbers: Smallest.ai: Rent a Phone Number $10/number/month on pay as you go, or bring your own carrier over SIP at cost; Indian phone numbers require regulatory compliance approval before activation, with registration and GST certificates (https://docs.smallest.ai/voice-agents/telephony/phone-numbers). Frequency: Indian (+91) numbers: buy a dedicated number through us, or connect your own Plivo number - Concurrent requests: Smallest.ai: Pay as you go lists 15 concurrent streams for text to speech, 100 concurrent streams for speech to text and a 450M chars/month cap; the docs limits page says 1 TTS request at a time per account; agents get 20 Included concurrent calls, Additional Concurrency: No (https://smallest.ai/pricing/models). Frequency: 5 on pay as you go, 25 on Business for Fourier TTS; a new workspace starts with 5 calls at a time, raised on request - Phone-line audio: Smallest.ai: PCM, WAV, MP3, or μ-law encoding, with sample rates from 8 kHz to 44 kHz; agents also take A-law at 8 kHz (https://docs.smallest.ai/models/text-to-speech/overview). Frequency: Raw 8 kHz μ-law or A-law; PCM or WAV, 48 kHz by default Checked 25 October 2026. https://scalefrequency.com/alternatives/smallest-ai ### Looking for a Murf alternative for phone calls? Murf charges less per character than we do, and it speaks more languages. Falcon is $0.01 per 1,000 characters against our ₹1.80. We spend the difference on voices built for phone calls in Hindi, Gujarati, Tamil, Telugu and English, five concurrent requests on pay as you go, GPUs in India on every plan, and full phone agents on the same key. - Published price: Murf: Falcon TTS API: $0.01 per 1,000 characters and Gen2 TTS API: $0.03 per 1,000 characters on pay as you go, after $10 worth of free credits; the API landing page says $0.01 / min instead (https://murf.ai/pricing?product=api). Frequency: ₹1.80 per 1,000 characters, on pay as you go and on Business (₹50,000/month commitment), before GST - Speech to text: Murf: Speech to Text API: $0.24 per hour, in Beta and India region only, with English (India) and 11 Indian languages and automatic language detection (https://murf.ai/pricing?product=api). Frequency: ₹18 per hour of audio, speaker labels included, charged once when the job completes - Hindi and Hinglish voices: Murf: Falcon 2.1's voice library lists 27 rows under Hindi - India, 8 under Hinglish - India, 5 under Tamil - India and 10 under Telugu - India; one voice can appear under several languages (https://murf.ai/api/docs/voices-styles/voice-library). Frequency: Four voices, neha, ashu, shruti and sameer, each speaking Hindi, English, Gujarati, Tamil and Telugu, with English words in Latin script - Languages: Murf: 150+ voices in 35 Languages, and Falcon 2.1 voices switch between Hindi and English inside one sentence (https://murf.ai/api). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish, in speech recognition and voice; Bengali, Marathi and Kannada are coming - Concurrent requests: Murf: Falcon 2.1 on the free tier and pay as you go: 5 for US-East; 2 for global and regional endpoints; Enterprise up to 10,000 (https://murf.ai/api/docs/resources/rate-limits). Frequency: 5 on pay as you go, 25 on Business; more on Enterprise - Where it runs: Murf: 11 regional environments for Falcon 2.1, India (in.api.murf.ai) among them, for Enterprise customers; account metadata stays in the United States; Speech to Text is India region only (https://murf.ai/api/docs/text-to-speech/data-residency). Frequency: Every call and every request runs on our own GPUs in India - Phone-line audio: Murf: HTTP streaming returns MP3, FLAC, WAV, ALAW, ULAW, OGG, or PCM; speech to text accepts 8 kHz µ-law and A-law (https://murf.ai/api/docs/text-to-speech/streaming). Frequency: Raw 8 kHz μ-law or A-law; PCM or WAV, 48 kHz by default - Telephony: Murf: No phone numbers, SIP trunking, campaigns or transfer in the API docs; Murf sells text to speech, speech to text, dubbing, translation, a voice changer and the Studio app (https://murf.ai/api). Frequency: Phone agents on the same key: ₹5 a minute pay as you go with the line included, +91 numbers from us or your own Plivo number - Cloning your own voice: Murf: Instant Voice Cloning is available on the Enterprise plan only, from one sample of up to 30 seconds; professional cloning takes 1 to 4 weeks (https://murf.ai/api/docs/voices-styles/instant-voice-cloning). Frequency: Custom voices: talk to us Checked 25 October 2026. https://scalefrequency.com/alternatives/murf ### Looking for an AssemblyAI alternative for Indian calls? AssemblyAI is a speech-to-text company with a published rate card, 99 languages on its older model and a free tier big enough to test with. We're narrower: recorded phone calls in India. Our transcript knows who spoke, keeps Hindi in Devanagari and English in Latin script, and costs ₹18 an hour with the speaker labels in. - Pre-recorded price: AssemblyAI: Universal-3.5 Pro $0.21 an hour and Universal-2 $0.15 an hour, pay as you go; speaker diarization is an add-on at $0.02 an hour on both (https://www.assemblyai.com/pricing). Frequency: ₹18 per hour of audio (₹0.30 a minute), charged once when the job completes; speaker labels included - Streaming price: AssemblyAI: Universal-3.6 Pro Realtime $0.45 an hour; Universal-Streaming $0.15 an hour, in English or six European languages; speaker diarization $0.12 an hour on streaming (https://www.assemblyai.com/pricing). Frequency: Streaming recognition runs inside every voice agent; as a standalone API it is in private preview, on request - Indian languages: AssemblyAI: Universal-3.5 Pro covers 18 languages with native code switching, Hindi the only Indian one; Universal-2 adds Bengali, Gujarati, Kannada, Malayalam, Marathi, Tamil and Telugu, which AssemblyAI's own page puts in its moderate (>25% to ≤50% WER) and fair (>50% WER) accuracy bands (https://www.assemblyai.com/docs/pre-recorded-audio/supported-languages). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish; Hindi in Devanagari, English in Latin script - Speaker labels: AssemblyAI: A paid add-on: $0.02 an hour on pre-recorded audio, $0.12 an hour on streaming (https://www.assemblyai.com/pricing). Frequency: Included on every transcript, with a speaker id per voice - An Indian voice: AssemblyAI: The Voice Agent API bundles AssemblyAI's own text-to-speech voices at $4.50/hr ($0.075/min), billed per second; the pricing page names no Indian-language voice (https://www.assemblyai.com/pricing). Frequency: Fourier TTS: Hindi, Gujarati, English, Tamil and Telugu, ₹1.80 per 1,000 characters - Where it runs: AssemblyAI: US endpoints in N. Virginia, Ohio and Oregon, EU endpoints in Frankfurt, Stockholm and Ireland; no India region (https://www.assemblyai.com/docs/pre-recorded-audio/select-the-region). Frequency: Every call and every request runs on our own GPUs in India - Try it free: AssemblyAI: No credit card; the free tier includes up to 185 hours of pre-recorded transcription and up to 333 hours of streaming transcription (https://www.assemblyai.com/pricing). Frequency: No sign-up to look around the sandbox; ₹200 of credit once you confirm a work email - Concurrent jobs: AssemblyAI: 5 parallel transcriptions on the free plan and 200+ on paid; streaming starts at 100 new sessions a minute on pay as you go (https://www.assemblyai.com/docs/pre-recorded-audio/rate-limits). Frequency: A shared endpoint; recordings up to 2 hours each Checked 25 October 2026. https://scalefrequency.com/alternatives/assemblyai ### Looking for a Speechmatics alternative for Indian calls? Speechmatics is a speech-to-text company with EU, US and Australian regions, on-prem containers for enterprises and a multilingual model that switches language mid-sentence. Its headline rate is reached only after two discounts. We transcribe Indian phone calls at one rate, ₹18 an hour, with speaker labels included and every word in its own script. - Pay-as-you-go price: Speechmatics: Base rates per hour of pre-recorded audio: Melia 1 $0.24, Standard 0.45, Enhanced 0.75 USD, billed to the second; the rate cards show $0.12, $0.23 and $0.38 after a 33% model-training opt-in and a 25% annual commitment (https://www.speechmatics.com/pricing). Frequency: ₹18 per hour of audio (₹0.30 a minute), charged once when the job completes; speaker labels included - Real-time price: Speechmatics: Base rates per hour of live audio: Linden 1 0.3 for voice agents, Melia 1 0.4, Standard 0.45, Enhanced 0.8 USD (https://www.speechmatics.com/pricing.md). Frequency: Streaming recognition runs inside every voice agent; as a standalone API it is in private preview, on request - Indian languages: Speechmatics: Bengali, Hindi, Marathi, Tamil and Urdu, plus a Tamil & English bilingual pack; no Telugu, Gujarati, Kannada or Malayalam in the table, and no Hindi and English pack (https://docs.speechmatics.com/speech-to-text/languages). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish; Hindi in Devanagari, English in Latin script - Speaker labels: Speechmatics: Included on pre-recorded audio at no extra charge; on real-time, Melia 1 does channel diarization only (https://docs.speechmatics.com/speech-to-text/models). Frequency: Included on every transcript, with a speaker id per voice - Where it runs: Speechmatics: EU1 (Europe), US1 (United States) and AU1 (Australia); no India region; on-prem containers, virtual appliances and on-device deployments on Enterprise only (https://docs.speechmatics.com/administration/regions). Frequency: Every call and every request runs on our own GPUs in India - Try it free: Speechmatics: Start with $100 in free credits, no card required; a volume discount applies above 500 hours a month (https://docs.speechmatics.com/administration/plans). Frequency: No sign-up to look around the sandbox; ₹200 of credit once you confirm a work email - Concurrent sessions: Speechmatics: 50 real-time sessions on self-serve (2 on Free) and 10 files a second pre-recorded; Enterprise adjusts to demand (https://www.speechmatics.com/pricing). Frequency: A shared endpoint; recordings up to 2 hours each Checked 25 October 2026. https://scalefrequency.com/alternatives/speechmatics ### Looking for a Whisper or gpt-transcribe alternative for Hindi calls? OpenAI's transcription API is cheap and everywhere: gpt-transcribe costs less than half a cent a minute, and most SDKs already speak it. But Whisper and the gpt-4o transcription models are being retired, the live model returns no speaker labels, and India is a storage region, not a processing one. We transcribe Indian phone calls at ₹18 an hour, with speaker labels included, on our own GPUs in India. - Transcription price: OpenAI speech-to-text: gpt-transcribe $0.0045 / minute; the deprecated gpt-4o-transcribe $0.006 / minute and gpt-4o-mini-transcribe $0.003 / minute until February 26, 2027 (https://developers.openai.com/api/docs/pricing). Frequency: ₹18 per hour of audio (₹0.30 a minute), charged once when the job completes; speaker labels included - Live transcription price: OpenAI speech-to-text: gpt-live-transcribe $0.017 / minute and gpt-realtime-whisper $0.017 / minute, over a Realtime session (https://developers.openai.com/api/docs/pricing). Frequency: Streaming recognition runs inside every voice agent; as a standalone API it is in private preview, on request - Whisper: OpenAI speech-to-text: whisper-1 is $0.006 a minute on its model page and deprecated, with removal from the API on February 26, 2027; OpenAI points to gpt-live-transcribe or gpt-transcribe instead (https://developers.openai.com/api/docs/deprecations). Frequency: One model, Fourier Transcribe, at ₹18 an hour - Indian languages: OpenAI speech-to-text: Whisper's own list has 98 languages, Hindi, Tamil, Telugu, Gujarati and Marathi among them, and OpenAI says accuracy varies by language; the docs print no language list for gpt-transcribe, which takes ISO language hints (https://developers.openai.com/api/docs/guides/speech-to-text). Frequency: Hindi, Gujarati, English, Tamil and Telugu, plus Hinglish; Hindi in Devanagari, English in Latin script - Speaker labels: OpenAI speech-to-text: Only on the deprecated gpt-4o-transcribe-diarize, until February 26, 2027; gpt-live-transcribe doesn’t return word-level timestamps, speaker labels, or transcription confidence scores (https://developers.openai.com/api/docs/guides/realtime-transcription). Frequency: Included on every transcript, with a speaker id per voice - Where it runs: OpenAI speech-to-text: India (in.api.openai.com) is on the data residency list for storage, not processing, and requires MAM or ZDR approval; residency endpoints carry a 10% uplift for newer models (https://developers.openai.com/api/docs/guides/your-data). Frequency: Every call and every request runs on our own GPUs in India - File limits: OpenAI speech-to-text: Files can be up to 25 MB; mp3, mp4, mpeg, mpga, m4a, wav and webm (https://developers.openai.com/api/docs/guides/speech-to-text). Frequency: Recordings up to 2 hours; files over 45 MB by https link - How to switch: OpenAI speech-to-text: POST /v1/audio/transcriptions with the file, and stream=true for partial text as it processes (https://developers.openai.com/api/docs/guides/speech-to-text). Frequency: POST /v2/transcriptions with the file, a line of context and the names to get right; for speech, the same OpenAI POST /v1/audio/speech protocol with model mira-tts Checked 25 October 2026. https://scalefrequency.com/alternatives/openai-stt ## Language pages ### Hindi Your customer picks up and hears Hindi. Everyday Hindi, said the way people say it on the phone. Four voices are ready today. Every call runs on our own GPUs, in India. - Phone agents: An agent calls your customers, or picks up when they call you. It confirms orders, reminds people about payments and books visits. - Text to speech: The same voices, as an API. Send Hindi text and get audio back. It costs ₹1.80 per 1,000 characters. - Speech to text: Send a recorded call and get a Hindi transcript. On the RinggAI Hindi benchmark it scored 9.12% normalised word error rate, the lowest of five systems. It trained on data from the same sources, and it isn’t first on every subset. - Write Hindi in Hindi: Send Hindi words in Devanagari. Romanised Hindi gets read with English vowels, and the caller hears an accent. Keep everyday Hindi. Formal Hindi sounds like a government notice, even when the audio is perfect. Write numbers as English words. The voice reads those most reliably. Voices: neha and shruti (female), ashu and sameer (male). Each has a real Fourier TTS sample on the page. https://scalefrequency.com/languages/hindi ### Hinglish Nobody on a phone call says भुगतान. They say payment. Our voices move between Hindi and English in the middle of a sentence, the way your customers do. One voice says the whole thing. - Phone agents: Most real calls in India are in Hinglish, so that’s what our agent templates speak. Order confirmations, reminders, follow-ups. - Text to speech: Send a mixed sentence and get one voice back. It costs ₹1.80 per 1,000 characters. - Speech to text: Transcripts come back the way people spoke. Hindi in Devanagari, English in Roman script, nothing translated. Common English words sometimes come back in Devanagari. - How the switching works: The model reads the script you send. Hindi words go in Devanagari. English words stay in Roman script, in the same sentence. Write numbers as English words, because the voice reads those most reliably. The console adds these rules to every agent template, so an agent’s replies come out mixed the right way. Voices: neha and shruti (female), ashu and sameer (male). Each has a real Fourier TTS sample on the page. https://scalefrequency.com/languages/hinglish ### Indian English Plenty of your customers would rather talk in English. Usually it’s the English they grew up with, not an American one. The same four voices that speak our Hindi speak English too. Every call runs on our own GPUs, in India. - Phone agents: Write the prompt in English and the agent calls in English. Same flows, same tools, same price as a Hindi call. - Text to speech: Send English text and get the same voices back. It costs ₹1.80 per 1,000 characters. - Speech to text: Recorded English calls transcribe too. Set the language to English, or let it detect. - Write it the way it should sound: The voice says what you write, so write numbers as words. Read IDs and pin codes digit by digit. Write the prompt in the language the call will be in. For English, that means an English prompt. Voices: neha and shruti (female), ashu and sameer (male). Each has a real Fourier TTS sample on the page. https://scalefrequency.com/languages/english ### Gujarati Your customer in Surat picks up and hears Gujarati. Not Hindi with one Gujarati word thrown in. Four voices speak it today. Every call runs on our own GPUs, in India. - Phone agents: An agent calls your customers in Gujarati, or picks up when they call you. Same flows, same tools, same price as a Hindi call. - Text to speech: Send Gujarati text to the same endpoint and get audio back. There is no language setting. It costs ₹1.80 per 1,000 characters. - Speech to text: Send a recorded call and get a Gujarati transcript, in Gujarati script. Nothing is translated. - Write Gujarati in Gujarati: Send Gujarati words in Gujarati script. That is what the voice reads. English words like order or payment stay in Roman script, in the same sentence, the way people say them. Write numbers as English words. The voice reads those most reliably. Voices: neha and shruti (female), ashu and sameer (male). Each has a real Fourier TTS sample on the page. https://scalefrequency.com/languages/gujarati ### Tamil Only 1.5% of Tamil speakers say they also speak Hindi. A Hindi call misses nearly all of them, however good it sounds. Our four voices speak Tamil now. Every call runs on our own GPUs, in India. - Phone agents: An agent calls your customers in Tamil, or picks up when they call you. Same flows, same tools, same price as a Hindi call. - Text to speech: Send Tamil text to the same endpoint and get audio back. There is no language setting. It costs ₹1.80 per 1,000 characters. - Speech to text: Send a recorded Tamil call and get the transcript back as it was spoken. Nothing is translated. - Write Tamil in Tamil script: Send Tamil words in Tamil script. That is what the voice reads. English words like order or payment stay in Roman script, in the same sentence, the way people say them. Write numbers as English words. The voice reads those most reliably. Voices: neha and shruti (female), ashu and sameer (male). Each has a real Fourier TTS sample on the page. https://scalefrequency.com/languages/tamil ### Telugu Your customer picks up and hears Telugu, said the way people say it on the phone. Only 5.7% of Telugu speakers also speak Hindi. This is the call that reaches the rest. Every call runs on our own GPUs, in India. - Phone agents: An agent calls your customers in Telugu, or picks up when they call you. Same flows, same tools, same price as a Hindi call. - Text to speech: Send Telugu text to the same endpoint and get audio back. There is no language setting. It costs ₹1.80 per 1,000 characters. - Speech to text: Send a recorded Telugu call and get the transcript back as it was spoken. Nothing is translated. - Write Telugu in Telugu script: Send Telugu words in Telugu script. That is what the voice reads. English words like order or payment stay in Roman script, in the same sentence, the way people say them. Write numbers as English words. The voice reads those most reliably. Voices: neha and shruti (female), ashu and sameer (male). Each has a real Fourier TTS sample on the page. https://scalefrequency.com/languages/telugu ## Tools ### AI calling cost calculator What your AI calls would cost a month, from ₹1 a minute or ₹5 pay as you go, and which plan is cheapest for you. https://scalefrequency.com/tools/ai-calling-cost-calculator ### Text-to-speech cost calculator What Fourier TTS would cost for your text, in characters or words, at ₹1.80 per 1,000 characters, and how many requests each plan runs at once. https://scalefrequency.com/tools/tts-cost-calculator ## Glossary ### 10-second billing blocks (pulse) A billing block, or pulse, is the unit a call's length is rounded up to before it's charged. With 10-second blocks, a call is billed in small slices instead of whole minutes. On a call: A COD confirmation call lasts 35 seconds. Billed per minute, you'd pay for a full minute. Billed in small blocks, you pay for 40 seconds. In Frequency: Every plan bills calls in 10-second blocks, rounded up, and you pay only when a call is answered. A call that rings out, finds the line busy or fails to connect costs nothing. Phone calling we supply can follow the carrier's own block, which the call record shows. https://scalefrequency.com/glossary/10-second-billing-blocks ### ASR vs STT ASR (automatic speech recognition) and STT (speech to text) are two names for the same job: turning spoken audio into written words. Researchers tend to say ASR, and API products tend to say STT. On a call: A caller says "मुझे refund चाहिए, order last week आया था". The ASR model, sold to you as an STT API, writes that sentence down so the agent's language model can read it and reply. In Frequency: Fourier STT is our own speech recognition, built for Indian phone calls in Hindi, English, Gujarati, Tamil, Telugu and Hinglish. Our agents listen with it, and recorded calls can go through our transcription API, which returns speakers and timestamps. https://scalefrequency.com/glossary/asr-vs-stt ### Backchannel A backchannel is a short sound a listener makes to show they're following, like "हाँ", "जी" or "hmm". It doesn't ask for the turn, so the speaker carries on. On a call: The agent is explaining an EMI due date and the caller says "हाँ हाँ, बोलिए" halfway through. That's a backchannel. The agent should finish its sentence instead of stopping to ask what the caller said. In Frequency: A caller's "haan", "ji" or "ok" never cuts our agents off, because we treat them as listening sounds. We've also built a listener that lets the agent say "हम्म" while the caller talks. It runs in shadow mode for now, logging where it would have spoken, until its timing holds up on real calls. https://scalefrequency.com/glossary/backchannel ### Barge-in Barge-in is when the caller starts talking while the voice agent is still speaking, and the agent stops to listen. Without it, people have to sit through every sentence before they can answer. On a call: The agent is reading out an order: "आपका order कल शाम तक…" The caller cuts in with "नहीं, मुझे cancel करना है." The agent stops mid-sentence and asks which order to cancel. In Frequency: You choose how much a caller has to say before the agent gives way: easy, normal, firm or never. A short "haan" or "ji" never stops it, while a clear "ruko" does at every setting except never. The agent also learns how loud the caller is, so a quieter voice in the room can't cut it off. https://scalefrequency.com/glossary/barge-in ### BYO telephony (BYOC) BYO telephony, or bring your own carrier (BYOC), means the voice AI platform makes and takes calls on your phone numbers and your telecom account. You keep the numbers, the carrier contract and the carrier's bill, and the platform supplies the agent. On a call: A clinic has printed one number on every prescription pad for years. With BYO telephony, the new AI receptionist answers that same number, and patients keep dialling the one they know. In Frequency: You can connect your own Plivo account and import a voice number you already own, for calls in and out on Indian +91 numbers. Calls on your own line carry no Mirai telco charge, and your provider bills you for them directly. If you'd rather not run a carrier account, rent a dedicated number through us instead. https://scalefrequency.com/glossary/byo-telephony ### Caller ID (CLI) and number pools Caller ID, or CLI (calling line identification), is the number a person sees when your call rings. A number pool spreads your outbound calls across several numbers, so no single one calls too many people too fast. On a call: A shopper misses your agent's call about her abandoned cart and rings the number back that evening. If the call came from your own number, your inbound agent picks up. If it came from a shared line, she reaches someone else's business. In Frequency: Pools spread outbound calls over your numbers at the hourly pace you set, and a contact's retries come from the number that rang them first. A number people stop answering gets rested before it's burnt, and a new one warms up at half pace for its first week. Calls from the Mirai shared line show a number other businesses use too, so callbacks to it won't reach you. https://scalefrequency.com/glossary/caller-id ### Caller-perceived latency Caller-perceived latency is the silence a caller hears between finishing their sentence and hearing the agent reply. It's the delay that counts, because it's the only one the caller experiences. On a call: The caller says "हाँ, कल दस बजे ठीक है" and stops. If the agent's "बढ़िया, मैंने book कर दिया" starts a moment later, the call flows. If it takes two seconds, the caller says "hello?" and both of them start talking at once. In Frequency: Every call runs on our own GPUs in India, with speech recognition, the model and the voice side by side. Our agents reach a first reply in a median of 900 ms, though real calls vary with the caller's network and how they speak. On the Realtime API, every turn reports its own voice-to-voice time, so you can watch our side of the wait. https://scalefrequency.com/glossary/caller-perceived-latency ### Character error rate (CER) Character error rate, or CER, scores a speech recogniser letter by letter instead of word by word. A small spelling slip costs a little, where word error rate would count the whole word as wrong. On a call: The caller said "पहुँचेगा" and the recogniser wrote "पहुंचेगा", with a dot where the chandrabindu should be. WER counts a whole wrong word. CER counts one character out of eight. In Frequency: We report CER next to WER for Fourier STT, so a reader can tell spelling noise from real misses. On the RinggAI Hindi benchmark its CER was 3.21%, with the full method and the test sets where we trail on the Fourier STT page. https://scalefrequency.com/glossary/character-error-rate ### Code-switching (Hinglish) Code-switching is moving between two languages inside one conversation, often inside one sentence. Hinglish, the mix of Hindi and English you hear on so many Indian business calls, is the everyday example. On a call: "आपका order confirm हो गया है, delivery कल evening तक होगी." That's one sentence, two languages. An agent built for pure Hindi stumbles on "confirm" and "evening"; one built for English can't say the rest. In Frequency: Our voices switch into English the way people really talk, so a Hinglish sentence comes out in one voice with no change of accent. Our transcripts write Hindi words in Devanagari and English words in Roman script, the way the call sounded. Hindi, English, Gujarati, Tamil and Telugu are live on every call, and so is the Hinglish mix. https://scalefrequency.com/glossary/code-switching ### Concurrency (calls and streams) Concurrency is how many calls, or speech streams, a system runs at the same moment. It's the number of lines open at once, and with call length it sets how many calls you can finish in an hour. On a call: A lender uploads a list of borrowers for EMI reminders. With five lines open and calls of under a minute, the list moves steadily all afternoon. With one line, the same list would run late into the evening. In Frequency: Calls over your concurrency limit wait in a queue and dial as soon as a line frees up, so going over the limit doesn't drop a call. The limit is a setting on our side, and raising it needs no change to your code. Text-to-speech plans each include a number of requests at once, listed on our pricing page. https://scalefrequency.com/glossary/concurrency ### DND and the NCPR (India) DND, short for Do Not Disturb, is the preference a mobile user in India registers with their telecom operator to block promotional calls and messages. It began as the National Customer Preference Register (NCPR); under TRAI's current rules the preferences sit on the operators' shared DLT platform. On a call: A customer once sent "BLOCK PROMO" to 1909. A campaign pitching them a new credit card is promotional, so calling them breaks their preference. A call about the delivery of something they ordered is a service call, and DND doesn't block it. In Frequency: We keep your own do-not-call list: every campaign checks it at dial time, so a number you add mid-campaign is skipped from that moment and never billed. We don't scrub the national DND register for you, so scrub your list before you upload it. https://scalefrequency.com/glossary/dnd ### Elo rating (Voice Arena) An Elo rating ranks voices by blind head-to-head votes. Listeners hear two clips of the same text without knowing who made them, pick the better one, and every vote nudges the two ratings apart. On a call: A listener hears two voices read "आपकी EMI इस शुक्रवार को due है" and picks the second because the first sounded flat on "शुक्रवार". Across thousands of such votes, the leaderboard settles into a ranking. In Frequency: Fourier TTS is on the Voice Arena Hindi leaderboard, where people pick between two clips without knowing who made them. Its rating was 993 when we last pulled the board, and the Fourier TTS page has the date, the margin, and where we lead and trail. https://scalefrequency.com/glossary/elo-rating ### End-of-turn detection End-of-turn detection predicts whether the caller has finished what they wanted to say and is handing the conversation back. It listens to the words and the voice as well as the silence. On a call: Caller: "मेरा order number है four five…" and a pause. The silence is long, but the words say more is coming. A good detector waits for the last two digits before the agent replies. In Frequency: On the Realtime API, server_vad and semantic_vad both use our own turn detection, so a client built for OpenAI's protocol works without retuning. Our agents also wait longer when the words say a number or an address isn't finished yet, and reply quickly after a plain yes or no. https://scalefrequency.com/glossary/end-of-turn-detection ### Endpointing Endpointing is how a speech system decides that the caller has stopped speaking. It is the moment the agent starts working on its reply. On a call: A customer reads out a pincode: "four zero zero… एक second… zero zero one." If endpointing fires at that pause, the agent hears half a pincode and asks for it again. In Frequency: Our agents wait longer while someone is reading out a number or an address, and answer sooner after a plain yes or no. On the Realtime API our turn detection makes this call for you, so the silence timers other providers expose are accepted and ignored. https://scalefrequency.com/glossary/endpointing ### μ-law and A-law (8 kHz phone audio) μ-law and A-law are the two standard ways phone networks squeeze voice into one byte per sample, defined in ITU-T G.711. A phone call carries 8,000 of those samples every second. On a call: Your voice bot's TTS returns crisp 48 kHz audio. The phone line wants 8 kHz μ-law. Ask the TTS for ulaw_8000 directly and you skip a resample and a conversion on every sentence of the call. In Frequency: Fourier TTS can send raw μ-law or A-law at 8 kHz, streamed as it's made, so the audio goes onto a phone line with no conversion on your side. Every response says its own rate and encoding in headers, so your code never has to guess. The Realtime API takes the same formats in and out, which saves you two resamples. https://scalefrequency.com/glossary/mu-law-a-law ### Node-based agent A node-based agent, or flow, runs a call as a map of steps. Each step has its own instructions, and the model moves to the next step when a condition in plain language is met. On a call: A loan reminder flow: confirm you're speaking to the borrower, state the EMI, then branch. "Wrong person" goes to a polite goodbye, "मैं कल pay कर दूँगा" goes to a step that reads the promised date back. In Frequency: A flow is one field on the agent, so it uses the same key, routes and billing as a prompt-based agent, on every plan. You can draft one from a plain-language brief or from a few recordings of good calls, then edit the nodes. Only the current node's instructions are in play, and the call keeps its full history as it moves. https://scalefrequency.com/glossary/node-based-agent ### Post-call analysis Post-call analysis reads a call after it ends and turns the conversation into something you can act on: a short summary, a yes or no against your goal, and the fields you asked for. On a call: After a COD confirmation call, the analysis returns: confirmed, yes; payment, online; new address, "Flat 4B, Sector 21"; send the payment-link WhatsApp template next. Your team never has to listen to the recording. In Frequency: Turn on analysis for an agent and every finished call gets read once more after the hangup, so the conversation itself never slows down. You get a summary, a success verdict against your own test and the fields from your schema, and the agent can even pick the WhatsApp or SMS template to send next. The result lands on the call and in the call.processed webhook. https://scalefrequency.com/glossary/post-call-analysis ### Prompt-based agent A prompt-based agent runs the whole call from one written brief: who it is, what it knows, and what it should get done. The language model decides each reply from that brief and the conversation so far. On a call: A clinic writes one prompt: you're Riya from the front desk, confirm {{patient_name}}'s appointment for {{date}}, offer to move it if needed, and never give medical advice. Every reminder call runs from that page. In Frequency: Every plan includes prompt-based and node-based agents, and the price is the same either way. A prompt can run up to 12,000 characters, with {{variables}} filled in for each call from your CSV or API request. Our prompting guide shows the shape we use: a role line, the facts, a short call flow and an example exchange. https://scalefrequency.com/glossary/prompt-based-agent ### SIP trunk A SIP trunk is a phone line delivered over the internet: a connection from a carrier that lets software place and receive ordinary phone calls. One trunk can carry many calls at once. On a call: A collections team pays a carrier for a trunk with a block of numbers. Its dialer sends a SIP INVITE down the trunk, the carrier rings the borrower's mobile, and the voice flows back and forth as packets until someone hangs up. In Frequency: Our docs list one way to bring your own line today: connect a Plivo account and import a +91 number you already own. A direct SIP trunk from your own carrier isn't on that list yet, so tell us if that's what you need. If you run your own SIP leg, Fourier TTS can still send audio straight into it as 8 kHz μ-law or A-law. https://scalefrequency.com/glossary/sip-trunk ### Speech-to-speech and realtime APIs A speech-to-speech API takes the caller's voice in and sends the agent's spoken reply back, on one live connection. You stream audio in and play audio out; listening, thinking and speaking all happen on the provider's side. On a call: Your Pipecat bot on a phone line streams the caller's "मुझे अपना plan बदलना है" to a realtime API. A moment later the reply comes back as audio, "ज़रूर, आप कौन सा plan लेना चाहेंगे?", and your code just plays it. In Frequency: Our Realtime API speaks OpenAI's Realtime protocol, so a client built for it only needs a new key and URL. Recognition, turn detection, the model and the voice run next to each other on one of our voice workers, and a session shows up in your call list like any other call. It's billed per minute in the same small blocks as a phone call, with no charge per token. https://scalefrequency.com/glossary/speech-to-speech-api ### SSML SSML, the Speech Synthesis Markup Language, is a W3C standard for adding instructions to text before a voice reads it. Tags can ask for a pause, stress a word, or say how a number or date should be read. On a call: An IVR prompt in SSML might read: आपका OTP है 4821 धन्यवाद. The tags ask for the code one digit at a time, then a half-second pause. In Frequency: Fourier TTS takes plain text. SSML isn't among the inputs our docs list, which are the text, the voice, the format and the sample rate. To change how something is read, write it the way it should sound, the same way our prompting guide writes amounts and identifiers. https://scalefrequency.com/glossary/ssml ### Streaming TTS Streaming text-to-speech sends audio back in small pieces while it's still being made, so playback can start before the whole sentence is ready. Live voice agents depend on it. On a call: The model writes "नमस्ते, मैं City Clinic से बोल रही हूँ, आपकी appointment कल है." With streaming, the caller hears "नमस्ते" while the rest is still being made. Without it, they sit through silence until the whole line is done. In Frequency: Our streaming endpoint keeps one WebSocket open for the whole call. You send text as your model writes it, we find the sentence ends, and the next sentence is made while the current one plays. When the caller interrupts, one message stops the audio that was playing. https://scalefrequency.com/glossary/streaming-tts ### TCCCPR and A2P voice calls TCCCPR is TRAI's rulebook for commercial calls and messages in India. An A2P (application-to-person) call is one that software places without a person dialling it, and TRAI's 2026 amendment gives it rules of its own. On a call: A lender's AI agent rings its borrowers with EMI reminders. Software dials every call and an artificial voice speaks, so these are A2P calls. Once the new rule takes effect, the lender has to declare them to its telecom operator first, along with the numbers they come from. In Frequency: Our docs are plain about who does what. We enforce the calling window you set on a campaign and your own opt-out list, while registration, consent and DND scrubbing stay with the business placing the calls. We don't sell 140-series numbers yet, and the console says so in one line when an outreach pool uses regular mobile numbers. https://scalefrequency.com/glossary/tcccpr-a2p-calls ### Text normalisation Text normalisation turns written shorthand into the words a voice should say. Digits, dates, currency amounts, abbreviations and codes all need it before text-to-speech can read them out. On a call: The text says "₹1,20,000 due on 15/07". Read naively, that's a string of digits and slashes. Normalised, it becomes "एक लाख बीस हज़ार रुपये, due on fifteen July", which is what a person would say. In Frequency: Fourier TTS reads your text as you send it, with no rewrite pass, so write amounts and dates the way they should sound. Our prompting guide has the rules we use for agents: amounts as words, and identifiers like pincodes read one digit at a time. https://scalefrequency.com/glossary/text-normalisation ### Time to first byte (TTS) Time to first byte, or TTFB, is how long a text-to-speech service takes to send back the first piece of audio after you ask for it. On a live call it decides how soon the caller hears the agent start speaking. On a call: The agent's model has written "जी, आपका order कल पहुँच जाएगा." The TTS request goes out. TTFB is the wait until the first slice of that sentence's audio arrives, which is the gap the caller hears before "जी". In Frequency: Every request in your Fourier TTS log shows first_byte_ms, so you can tell our share of a wait from your network's. In our load test with many streams on one GPU, nearly every request got its first audio within 275 ms. That's measured next to the engine, so your own network time comes on top. https://scalefrequency.com/glossary/time-to-first-byte ### Tool calling in voice agents Tool calling lets a voice agent reach your systems during a conversation, such as checking an order or booking a slot. The model decides when it needs a tool, asks for it with the right inputs, and uses the answer in its reply. On a call: Caller: "मेरा parcel कहाँ है?" The agent calls your order-status API with the phone number it already has, gets "out for delivery", and says "आज शाम तक पहुँच जाएगा" without putting the caller on hold. In Frequency: Tools run before a call to fetch context, during it when the agent decides it needs one, and after it to send the outcome to your CRM. An on-call tool can make the agent wait for the answer, or run in the background and hand the result back when it's ready. Secrets stay in the workspace and never come back to the editor. https://scalefrequency.com/glossary/tool-calling ### Turn-taking Turn-taking is how people in a conversation share the floor: one speaks, then the other, with short gaps and little overlap. A voice agent has to do it as well as a person, or the call feels off within seconds. On a call: Agent: "Delivery कल सुबह ठीक रहेगा?" Caller: "हाँ… नहीं, शाम को." A human hears that the turn isn't over after "हाँ". An agent with poor turn-taking replies to the yes and has to walk it back. In Frequency: When we change how our agents take turns, we test it on scripted calls first and count what goes wrong: dead air over a second and a half, restarts, repeated lines, sentences cut in half. Blind judges and a person with headphones then compare the new build with the old one before we merge it. https://scalefrequency.com/glossary/turn-taking ### Voice activity detection (VAD) Voice activity detection, or VAD, is the part of a voice system that tells speech apart from silence and background noise, many times a second. Everything else in the call waits on its answer. On a call: A delivery customer takes the call at a busy signal, with a horn going off every few seconds. A weak VAD hears the horn as speech, keeps the turn open, and the agent never replies. In Frequency: A cough or a hum on the line used to be able to cancel a reply our agent had already started. Now a sound with no words in it can't: the agent picks the reply back up and carries on. With adaptive interruptions on, the agent also learns the caller's own volume, so quieter voices in the room are ignored. https://scalefrequency.com/glossary/voice-activity-detection ### Webhook (call events) A webhook is a message the calling platform sends to your server the moment something happens on a call: it started, it ended, it went to voicemail, or its analysis is ready. Your system reacts without having to keep asking. On a call: A COD confirmation call ends at 4:12 pm. Your server gets call.completed right away, then call.processed with "confirmed: yes" once the analysis is done, and your warehouse system releases the parcel for packing. In Frequency: Pass a webhook URL when you create a call and we send a signed event at each step, with exactly one ending event per call. Each delivery carries an HMAC-SHA256 signature over the raw body, and if your server is down we keep retrying for about 36 minutes. https://scalefrequency.com/glossary/webhook ### Word error rate (WER) Word error rate, or WER, measures how often a speech recogniser gets words wrong. It counts the words it swapped, missed or added, and divides by the number of words that were really said. On a call: The caller said "मुझे कल शाम को delivery चाहिए". The recogniser wrote "मुझे कल शाम delivery चाहिए" and dropped "को". That's one word missing out of six, so the WER for this sentence is about 17%. In Frequency: On the public RinggAI Hindi benchmark, our recogniser scored 9.12% normalised WER, the lowest in our report. It's one benchmark, and an in-domain one for us, so the Fourier STT page also shows the test sets where we trail. https://scalefrequency.com/glossary/word-error-rate ## Journal - [How we decide a voice agent sounds more natural](https://scalefrequency.com/journal/how-we-decide-a-voice-agent-sounds-natural): Main beat or matched our changes for three rounds. Here's what finally moved in the fourth, which numbers we trusted, and the evidence we threw out. - [Releasing a voice stack without dropping live calls](https://scalefrequency.com/journal/releasing-a-voice-stack-without-dropping-calls): A phone call holds one process for minutes. So a release can't restart anything a call is using, and it has to be able to change its mind halfway. - [Our cloud vendor went dark for six hours. Our customers never knew.](https://scalefrequency.com/journal/our-cloud-vendor-went-dark): The failure mode was not the outage. It would have been our own controller reacting to it. - [The deploy log is the git log](https://scalefrequency.com/journal/the-deploy-log-is-the-git-log): A change with no commit is a change with no history, no review and no way back. So we removed the ability to make one. - [We ship voice by measurement, not by vibes](https://scalefrequency.com/journal/we-ship-voice-by-measurement): The ear listens to one sample. The defect lived between samples. Pitch analysis found it, and we rolled back the same hour. ## Privacy Optional Google Analytics loads only with permission. Browser demo audio and transcripts may be stored. The calendar loads only when opened. See the full privacy notice for provider and data-handling details. https://scalefrequency.com/privacy Privacy contact: sneh@miraiminds.co ## Terms The sandbox, console and API are covered by our terms: acceptable use, how calls are processed (our own GPUs in India, with named backup providers), what we keep and for how long, payment and liability. https://scalefrequency.com/terms