Drop in to existing OpenAI code

Already using the OpenAI Python SDK? Keep your code. Point base_url to Shunya and you instantly gain 23 Indic languages, 46 voices, and on-prem deployment, no rewrites, no new client library.

export ACCESS_TOKEN="eyJhbGciOiJSUzI1NiIs…paste-here"

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

Your journey

Step 1: The two-line swap

Take any working OpenAI TTS call. Change two lines, the base_url and the voice: and you're calling Shunya:

Before: vanilla OpenAI

from openai import OpenAI

client = OpenAI(api_key="sk-openai-...")

response = client.audio.speech.create(
    model="tts-1",
    input="Hello, how are you today?",
    voice="alloy",
    response_format="mp3",
)
response.stream_to_file("output.mp3")

After: same SDK, Shunya backend

from openai import OpenAI
import os, requests

# swap 0 (auth): the speech APIs take a short-lived token, not the raw key — mint one, pass as api_key
token = requests.post("https://app.shunyalabs.ai/api/auth/token",
                      headers={"api-key": os.environ["SHUNYALABS_API_KEY"]}).json()["token"]

client = OpenAI(
    api_key=token,               # ← swap 0 (a minted token, refresh as it expires)
    base_url="https://ttsv2.shunyalabs.ai/v1",   # ← swap 1
)

response = client.audio.speech.create(
    model="zero-indic",                         # ← swap 2 (model)
    input="Hello, how are you today?",
    voice="Varun",                              # ← swap 3 (voice)
    response_format="mp3",
)
response.stream_to_file("output.mp3")

Step 2: What stays the same

• Client class, OpenAI / AsyncOpenAI from the official openai package.
• Method shape, client.audio.speech.create(...), same fields.
• Response shape, response.stream_to_file(), response.read(), response.iter_bytes() all work as in OpenAI.
• Auth pattern, Bearer token via api_key; the SDK sets the Authorization header for you.
• response_format, mp3, wav, pcm, flac, ogg_opus all map across.
• Error handling, openai.AuthenticationError, openai.RateLimitError, openai.APIError behave as expected.

Step 3: What changes

Step 4: Voice mapping cheatsheet

OpenAI ships generic English voices. Shunya gives you native Indic voices that can also speak English (cross-lingual). Pick by language and gender:

See the full voice list for all 23 languages × 2 genders = 46 voices, plus details on cross-lingual synthesis (any voice, any language).

Step 5: Async + streaming with the OpenAI SDK

from openai import AsyncOpenAI

client = AsyncOpenAI(
    api_key=os.environ["ACCESS_TOKEN"],
    base_url="https://ttsv2.shunyalabs.ai/v1",
)

async with client.audio.speech.with_streaming_response.create(
    model="zero-indic",
    input="Hello, how are you today?",
    voice="Sunita",
    response_format="pcm",
) as response:
    async for chunk in response.iter_bytes():
        play(chunk)   # bytes arrive as the model generates them

Step 6: Pipeline: GPT-4 → Shunya TTS

Use one OpenAI client for the LLM, a second OpenAI client (pointed at Shunya) for TTS. Stream tokens, flush at sentence boundaries:

from openai import AsyncOpenAI

oai     = AsyncOpenAI()                                      # default OpenAI base_url
shunya  = AsyncOpenAI(base_url="https://ttsv2.shunyalabs.ai/v1",
                      api_key=os.environ["ACCESS_TOKEN"])

buffer = ""
stream = await oai.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Tell me a story"}],
    stream=True,
)
async for chunk in stream:
    token = chunk.choices[0].delta.content or ""
    buffer += token
    if token in (".", "!", "?", ";") and len(buffer) > 15:
        async with shunya.audio.speech.with_streaming_response.create(
            model="zero-indic", input=buffer, voice="Sunita", response_format="pcm",
        ) as resp:
            async for audio in resp.iter_bytes():
                play(audio)
        buffer = ""

Step 7: Why teams choose this path