Use Fourier TTS with an OpenAI-compatible speech client.

Only the base URL, the model and the voice change.

The base URL, the model and the voice. Keep the key on your server, never in a browser or mobile app.

Setting	Mira value
Base URL	https://sandbox.voice.miraiminds.co/v1
Request path	/audio/speech, under that base URL
Model	mira-tts (the older mira-tts-v51 still works as an alias)
Voice	ashu, neha, shruti or sameer; neha if you leave it out
Output	Always set response_format
Input	Up to 2,000 characters

Fourier TTS costs ₹1.80 per 1,000 characters. Building a whole phone agent? Those cost ₹1 a minute on our volume plan; see AI calling.

For a key, open Developers in the sandbox and press Create key. The sk_live_ key shows once and works across the v2 API. The same handler also sits at POST /v2/tts, which takes text in place of input.

Your first request saves a WAV file.

This saves a WAV file and spends a few paise. Put your key in MIRA_API_KEY first.

curl --fail-with-body \
  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।"
  }' \
  --output speech.wav

If it fails, --fail-with-body writes the error into speech.wav, so check before you press play.

Pick the format and sample rate you need.

response_format	You get	Streams?
pcm	Raw signed 16-bit little-endian mono, no header	Yes
mulaw (also ulaw, pcm_mulaw)	Raw G.711 μ-law, one byte per sample	Yes
alaw (also pcm_alaw)	Raw G.711 A-law, one byte per sample	Yes
wav	A playable 16-bit mono file	No, it comes when synthesis ends

sample_rate takes 8000, 16000, 22050, 24000, 44100 or 48000, with μ-law and A-law at 8000 or 16000 only. Defaults are 48000 for pcm and wav, 8000 for μ-law and A-law. A rate in the format, like pcm_16000, wins; a conflicting sample_rate gets a 400.

On a phone line, ask for the line’s format and forward chunks as they come. Most lines carry 8 kHz μ-law, so ulaw_8000 needs no decoding or resampling.

You can stream with the OpenAI Python SDK.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MIRA_API_KEY"],
    base_url="https://sandbox.voice.miraiminds.co/v1",
)

with client.audio.speech.with_streaming_response.create(
    model="mira-tts",
    voice="ashu",
    response_format="pcm",
    input="चलिए, आपका order कल deliver होगा।",
    extra_body={"sample_rate": 16000},
) as response:
    print(response.headers.get("x-sample-rate"), response.headers.get("x-audio-encoding"))
    with open("speech_16k.pcm", "wb") as out:
        for chunk in response.iter_bytes():
            out.write(chunk)  # raw s16le mono at 16 kHz; feed a player here instead

We ran this with openai 1.109.1 against the sandbox on 16 October 2026. It printed 16000 pcm_s16le and wrote 16 kHz raw PCM.

Play chunks as they arrive, or there’s no point streaming. For live conversations, the WebSocket endpoint is better: one connection per call, text as your model writes it, and one message to stop when the caller interrupts.

The headers tell you how to play the audio.

Every 200 carries these:

Header	Meaning
X-Request-Id	Request ID (tts_…), as in your request log
X-Chars-Billed	Characters billed
X-Cost-Paise	Cost, once it completes
X-Sample-Rate	Sample rate you got
X-Audio-Encoding	pcm_s16le, mulaw or alaw
X-Mira-Model	The build that made it, like miraiminds/mira-tts-v53

Raw audio has no header, so read X-Sample-Rate and X-Audio-Encoding before you play it. Get them wrong and you hear noise, or the wrong speed. Note X-Mira-Model when you compare releases. GET /v2/tts/requests lists every request with first_byte_ms, total_ms and cost.

Existing adapters often trip on OpenAI’s defaults.

Adapters often bake in OpenAI’s defaults. Check yours doesn’t:

• reject neha or shruti as unknown voices,
• ask for MP3 unless you override it,
• assume a rate, like 24 kHz, instead of reading X-Sample-Rate,
• buffer the whole response even though we stream it.

On Pipecat, use pipecat-mirai instead of the OpenAITTSService workaround; see the Pipecat guide. Anywhere else, listen to what actually plays.

These are the errors you might see, and the fixes.

Errors carry X-Request-Id. Quote it when you ask us for help.

• 400 invalid_request: fix the field it names, such as empty input, an unknown voice or a format we don’t do.
• 401 unauthorized: check the Authorization: Bearer sk_live_… header.
• 402 insufficient_balance: top up. Nothing was made or charged.
• 404 model_not_found: use mira-tts.
• 429 at_capacity or rate_limited: wait for Retry-After, then resend.
• 502 upstream_error: retry. 503 fleet_offline: wait for Retry-After. Neither is charged.

Cap retries on a live call: a sentence that plays after the moment has passed is worse than none. Our request log keeps only the first 48 characters of each input. The latency guide helps you tell a slow request from a playback problem.