Choosing a Hindi and Hinglish text-to-speech API.

What should I test?

The lines your product will actually say. A fluent greeting tells you very little. Names, addresses, English acronyms and amounts are where a voice works for you or doesn’t.

Write Hindi in Devanagari and English words in Latin script, the way people type Hinglish: आपका order कल deliver होगा। Romanised Hindi like aap kab tak kar payenge gets read with English vowels. Spell out amounts and dates as words; the prompting guide explains why.

Some sample lines to start with, not measured results:

Test	Example input	Listen for
Plain Hindi	नमस्ते, आपका ऑर्डर तैयार है।	Natural phrasing and a clean ending
Code-switching	आपका refund process हो गया है।	Smooth switches between languages
A correction	माफ़ कीजिए, delivery address बदलना है।	Emphasis without the voice changing character
An acronym	आपका OTP भेज दिया गया है।	Whether the abbreviation is clear
An amount	आपकी EMI पाँच हज़ार दो सौ रुपये है।	Whether the number is easy to follow
Your own term	A product or brand name from your catalogue	A pronunciation your users recognise

Ask native speakers to describe each error. “Sounds unnatural” gives an engineer nothing to fix. “Stresses the second syllable of the brand name” does.

Building a whole phone agent instead? Agents cost ₹1 a minute; see AI calling.

How do I make my first request?

Create a key under Developers in the sandbox, no access code needed, and send one request:

curl https://sandbox.voice.miraiminds.co/v1/audio/speech \
  -H "Authorization: Bearer $MIRA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mira-tts",
    "voice": "neha",
    "response_format": "wav",
    "input": "नमस्ते, आपका ऑर्डर तैयार है। आपका refund process हो गया है।"
  }' \
  --output speech.wav

On 16 October 2026 this returned a 48 kHz mono 16-bit WAV, billed 59 characters, and reported X-Mira-Model: miraiminds/mira-tts-v53. The OpenAI-compatible setup guide covers other formats, headers and errors.

Keep your test lines in your repository, and save X-Mira-Model with every result, so you can tell a model update from a change in your text.

How are the four voices different?

Every plan has the same four: ashu and sameer are male, neha and shruti female, and neha is the default. All four are Hindi voices that switch into English. GET /v2/voices lists them without audio, so hear them in Agent Studio’s voice picker. An unknown voice gets a 400 naming the real ones; we never swap in another quietly.

Listen across a run of turns. A voice can sound fine in every clip and still drift in pitch from one turn to the next. Our post on measuring voice releases tells how that drift caught a regression our ears missed.

Does streaming change anything?

Yes. pcm, mulaw and alaw stream, so the first bytes arrive while the rest is still being made. wav arrives in one piece at the end, which suits files, not conversations.

On a phone call, ask for the line’s own format, usually ulaw_8000, and forward chunks as they come. 48 kHz PCM needs 768 kbit/s per stream; 8 kHz μ-law needs 64. For a whole conversation, the WebSocket endpoint keeps one connection open, takes text as your model writes it, and finds sentence ends itself, the danda (।) included.

A saved file doesn’t prove live playback works; the latency guide shows how to time what the listener hears. On Pipecat, the Pipecat guide handles streaming, barge-in and phone pacing.

What about other Indian languages?

Today the engine speaks Hindi, English, Gujarati, Tamil and Telugu, and the Hinglish in between, with the same four voices. Send each language in its own script; there is no language field to set. Bengali, Marathi and Kannada are on the language roadmap, each dated when it passes evaluation. Hear the voices on the language pages.

Check your language on the endpoint you’ll use before a rollout. To hear the voices in use, visit the text-to-speech model page.