Text-to-Speech (TTS)

Radiance is Shunya's speech synthesis family. 46 speaker voices across 23 Indic languages and English. Every voice can speak every language. 11 expression styles. Voice cloning from a 3-6 second reference clip.

First: get an access token

The examples below send Authorization: Bearer $ACCESS_TOKEN. The speech APIs accept only a short-lived access token — never your API key directly.

From the Shunya Playground (recommended). Open API keys in the Playground, click Generate token next to your API key, and copy it — then set it:

shell

export ACCESS_TOKEN="eyJhbGciOiJSUzI1NiIs…paste-here"

Or mint it from your API key — the path for production, where your app refreshes the token as it nears expiry (the response carries expires_in):

shell

export ACCESS_TOKEN=$(curl -s -X POST https://app.shunyalabs.ai/api/auth/token \
  -H "api-key: $SHUNYALABS_API_KEY" | jq -r .token)

Radiance models

Pass the model id in the model field. The product names are for reading; the ids are what the API accepts (the earlier zero-* ids still work).

Model	Model id	Languages	Use it for
Radiance Indic	`radiance-indic`	23 Indic languages + English	The default. All 46 named voices, expression styles and voice cloning.
Radiance Oriental	`radiance-oriental`	Japanese, Korean	Native Japanese and Korean speech.
Radiance Universal	`radiance-universal`	Extended world languages (see Voices & languages)	French, German, Spanish, Arabic and other non-Indic languages.

How it fits together

Batch vs Streaming

Two synthesis modes are available. Same model, same voices, different transport and different "when does the first byte of audio leave the server."

Batch

HTTP POST, returns a complete file

Send text via HTTP POST and receive a complete audio file in a single response.

Best for

Pre-rendered voice prompts for IVR and telephony systems.
Notification audio, order updates, alerts, reminders.
Podcast, audiobook, and long-form content generation.
Any use case where audio does not need to start playing before synthesis is complete.

Properties

Transport: HTTP POST
Endpoint: https://ttsv2.shunyalabs.ai/v1/audio/speech
Auth: Bearer <ACCESS_TOKEN>
Required: text, model, voice
Default format: mp3

Flow

1POST text→2Server synthesizes→3Receive audio

Streaming

WebSocket, chunks arrive in real time

Open a persistent WebSocket connection and receive audio chunks in real time as synthesis happens.

Best for

Voice agents and conversational AI requiring sub-second audio start.
IVR and telephony pipelines.
Real-time audio playback in applications.
Any use case where audio must begin playing before synthesis of the full text is complete.

Properties

Transport: WebSocket
Endpoint: wss://ttsv2.shunyalabs.ai/v1/realtime
also: /ws/tts, /ws
Config: TTSConfig
Default format: mp3

Lifecycle

1Connect→2Receive chunks→3Done

Source: Shunyalabs TTS Developer Documentation v1.0 (March 2026), §2.1 Batch overview and §3.1 Streaming overview, text reproduced verbatim.

Endpoints

Mode	Endpoint	Default format
Batch	`POST https://ttsv2.shunyalabs.ai/v1/audio/speech`	mp3
Streaming	`wss://ttsv2.shunyalabs.ai/v1/realtime`	mp3 (pcm recommended)
Health	`GET https://ttsv2.shunyalabs.ai/health`	-

Your first synthesis

shell

curl -X POST https://ttsv2.shunyalabs.ai/v1/audio/speech \
  -H "Authorization: Bearer $ACCESS_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model":"radiance-indic","input":"Hello, how are you today?","voice":"Varun"}' \
  --output hello.mp3

python

import asyncio
from shunyalabs import AsyncShunyaClient
from shunyalabs.tts import TTSConfig

async def main():
    async with AsyncShunyaClient() as client:
        result = await client.tts.synthesize(
            "Hello, how are you today?",
            config=TTSConfig(model="radiance-indic", voice="Varun"),
        )
        result.save("hello.mp3")

asyncio.run(main())

python

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["ACCESS_TOKEN"],
    base_url="https://ttsv2.shunyalabs.ai/v1",
)
response = client.audio.speech.create(
    model="radiance-indic",
    input="Hello, how are you today?",
    voice="Varun",
    response_format="mp3",
)
response.stream_to_file("output.mp3")

Key features

46 voices

Male & female speakers per language. Any voice, any language, cross-lingual synthesis is built in.

7 audio formats

PCM, WAV, MP3, OGG Opus, FLAC, mulaw, alaw. Pick by use case, telephony, web, archival.

11 expression styles

Happy, Sad, News, Narrative, Conversational, Enthusiastic, and more, prepended as tags to your text.

Voice cloning

Clone a voice from 3-6 seconds of reference audio. Works across all 23 supported languages.

Streaming

First audio in under 350 ms. Critical for voice agents.

LLM → TTS pipeline

Pipe OpenAI / Anthropic / Gemini tokens to TTS at sentence boundaries, the core pattern for low-latency voice agents.

Required fields, at a glance

json

{
  "model": "radiance-indic",         // required
  "input": "Your text here",     // required, up to 10,000 chars
  "voice": "Varun",              // required, see Voices page for full list
  "response_format": "mp3",      // optional, default mp3
  "speed": 1.0,                  // optional, 0.25 to 4.0
  "language": "en",              // optional, ISO code for preprocessing
  "trim_silence": false,         // optional, tight audio when true
  "quality": "low",              // optional, low | medium | high
  "latex": false,                // optional, speak LaTeX math aloud
  "background_audio": null,      // optional, ambient preset or base64
  "background_volume": 0.1,      // optional, 0.0 to 1.0
  "reference_wav": "...",        // optional, base64 for voice cloning
  "reference_text": "..."        // optional, transcript for voice cloning
}