Text To Speech

Text-to-Speech — Bulbul

Live in-chat TTS → sarvam-mcp (sarvam_tools_tts_*). This skill = SDK code.

[!IMPORTANT] Auth: api-subscription-key header — NOT Authorization: Bearer. Base URL: https://api.sarvam.ai (NOT /v1 — that prefix is only for the OpenAI-compatible chat endpoint) SDK floor for the code below: Python sarvamai>=0.1.29, JS sarvamai@>=1.1.8. On older versions the TTS param is target_language_code — see Gotchas.

Model

bulbul:v4-flash — latest, use this for new integrations. Same REST / HTTP stream / WebSocket contract as v3. 200+ persona voices (222 IDs).

bulbul:v3 — what the API uses when model is omitted. Short names (shubh, priya). Still the right model for Malayalam and Odia, which have no v4 personas yet.

Speakers (v4 Flash)

IDs are lowercase voice_language_style: aparna_hi_customer, simran_en_customer, shubh_enhi_ads. Omit speaker and the default is shubh_enhi_ads. v2/v3 short names (shubh, priya, ishita, anushka) return 400.

The middle token is a language tag (hi, en, bn, gu, kn, mr, pa, ta, te) or enhi for English–Hindi code-mix. Use an _enhi_ persona when one utterance is genuinely mixed; a few English loanwords in Hindi script do not need it.

language_code still takes the usual BCP-47 values (hi-IN, en-IN, … including ml-IN and od-IN). There are no ml or od persona IDs — for those languages set model="bulbul:v3" and a v3 short name.

Do not copy speaker="ishita" from the Voices “generate your own sample” snippet — that is a v3 short name and fails on v4-flash.

Coverage by language (live-verified 2026-10-06; one working persona each — pick others from the live list):

The roster changes. Sarvam renamed voices on 2026-10-06 (bappa_bn_conversation → arnab_bn_conversation, vetri_ta_ads/vetri_ta_suspense → aravind_ta_*), and the old IDs now 400. Do not hardcode IDs from memory or old blog posts. An unknown speaker returns 400 with the full live list, so the list is one bad request away. The API also lists 2 Assamese (_as_) personas that the docs catalog does not.

Quick Start (Python)

pythonCopy code
from sarvamai import SarvamAI
from sarvamai.play import save

client = SarvamAI()

response = client.text_to_speech.convert(
    text="नमस्ते, आप कैसे हैं?",
    language_code="hi-IN",
    model="bulbul:v4-flash",
    speaker="aparna_hi_customer",
    enable_preprocessing=True,
)
save(response, "output.wav")

# HTTP Stream (lower latency, binary audio)
chunks = []
for chunk in client.text_to_speech.convert_stream(
    text="Hello from Sarvam AI",
    language_code="en-IN",
    model="bulbul:v4-flash",
    speaker="simran_en_customer",
    output_audio_codec="wav",
):
    chunks.append(chunk)
audio = b"".join(chunks)

Quick Start (JavaScript/TypeScript)

typescriptCopy code
import { SarvamAIClient } from "sarvamai";
import { writeFile } from "fs/promises";

const client = new SarvamAIClient({ apiSubscriptionKey: "YOUR_SARVAM_API_KEY" });

// REST
const response = await client.textToSpeech.convert({
    text: "नमस्ते, आप कैसे हैं?",
    language_code: "hi-IN",
    model: "bulbul:v4-flash",
    speaker: "aparna_hi_customer",
    enable_preprocessing: true,
});

// HTTP Stream (lower latency, returns BinaryResponse)
const streamResponse = await client.textToSpeech.convertStream({
    text: "Hello from Sarvam AI",
    language_code: "en-IN",
    model: "bulbul:v4-flash",
    speaker: "simran_en_customer",
    output_audio_codec: "wav",
});
const bytes = await streamResponse.bytes();
await writeFile("output.wav", bytes);

WebSocket Streaming

pythonCopy code
import asyncio
from sarvamai import AsyncSarvamAI

async def tts_stream():
    client = AsyncSarvamAI()
    async with client.text_to_speech_streaming.connect(model="bulbul:v4-flash") as ws:
        await ws.configure(
            target_language_code="hi-IN",
            speaker="aparna_hi_customer",
            output_audio_codec="wav",
        )
        await ws.convert("Your text here")
        await ws.flush()
        async for message in ws:
            pass  # base64 audio chunks

asyncio.run(tts_stream())

Character Limits

Gotchas

Full Docs

Fetch the voice catalog, ranked picks, streaming protocol, and codec options from:

• https://docs.sarvam.ai/llms.txt — comprehensive docs index
• Bulbul models
• Voices (v4 catalog + v3 previews)
• Bulbul v4 Flash best practices
• TTS Overview
• HTTP Stream
• Pronunciation Dictionary
• Rate Limits