Choosing a Hindi and Hinglish text-to-speech API

What makes a Hindi TTS evaluation useful?

Test the speech your product will generate. A fluent greeting tells you little. Customer names, local addresses, English acronyms and amounts are where a voice either works for your application or doesn’t.

Write Hindi in Devanagari and leave English words in Latin script, as people type Hinglish: आपका order कल deliver होगा। Romanised Hindi such as aap kab tak kar payenge gets read with English vowels. The prompting guide also recommends spelling out amounts and dates in words, because digits and currency symbols can be read inconsistently.

Mira TTS is one of Mirai’s model APIs, alongside speech recognition, a language model and voice agents, and it costs ₹1.80 per 1,000 characters. Building a full phone agent instead? Agents start at ₹1 a minute at volume; see AI calling.

These are sample test lines to start a set with. They are 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 they hear. “Sounds unnatural” doesn’t tell an engineer what to change; “stresses the second syllable of the brand name” does.

How do you generate the first sample?

Create a key under Developers in the sandbox. No access code is needed. Then 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

We ran this request against the sandbox on 16 October 2026. It 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 the other formats, the response headers and the errors.

Keep a small evaluation set in your repository: the exact input, pronunciation notes, the voice, and why each line is there. Save the X-Mira-Model value with each result so you can tell a model update apart from a change in your own text. Change one thing at a time (model, voice or text formatting) so you can tell what caused a difference.

How do the four voices differ?

The catalogue is the same on every plan: ashu and sameer are male, neha and shruti are female, and neha is the default. All four are Hindi voices that code-switch into English. GET /v2/voices returns the live list; it has no audio samples, so use the voice picker in Agent Studio to hear each one. An unknown voice returns a 400 that names the allowed set. The API never swaps in another voice silently.

Listen to each candidate across a sequence of turns. A voice can sound fine in each clip and still drift in pitch or delivery from one turn to the next. Mirai’s voice measurement journal describes a release where turn-to-turn pitch spread caught a regression that listening to single samples missed.

How does streaming change the integration?

The format decides whether you can start playback early. pcm, mulaw and alaw stream: the first bytes arrive while the rest is still being synthesised. wav arrives in one piece once synthesis finishes, which is fine for files and wrong for a live conversation.

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

Saving a valid file does not prove live playback works. Measure the delay before the listener hears speech separately from the API’s first-byte time; the latency guide explains how. If you build on Pipecat, the Pipecat guide handles streaming, barge-in and phone pacing for you.

What about other Indian languages?

The engine speaks Hindi, English and Hinglish today. More Indian languages are on the language roadmap, with each date announced as the language clears evaluation. Speech synthesis ships first, and full agent support in a language follows once recognition and the model reach the same bar.

Check the status of the exact language you need, on the endpoint you plan to use, before you commit to a rollout. Then run the same listening test for it that you ran for Hindi. To see the voices in use, visit the text-to-speech model page.