Sarvam TTS plugin guide

Overview

Use the Sarvam TTS plugin to synthesize Indian-language and English speech in LiveKit Agents. It provides natural Indic voices, low-latency turn-taking, configurable speaking style, and production audio formats for browser, mobile, and telephony use cases.

For new voice agents, start with bulbul:v3, set target_language_code explicitly, and choose a speaker that is compatible with the selected model.

Authentication

The Sarvam plugin requires a Sarvam API key .

Set SARVAM_API_KEY in your .env file:

SARVAM_API_KEY=<your-sarvam-api-key>

Installation

Install the plugin:

uv add "livekit-agents[sarvam]~=1.5"

pnpm add @livekit/agents-plugin-sarvam@1.x

Usage

Use Sarvam TTS within an AgentSession or as a standalone speech generator. For example, you can use this TTS in the Voice AI quickstart.

For most LiveKit voice agents, begin with the following settings. Explicit configuration makes voice quality, latency, and deployment behavior easier to reproduce across environments.

• target_language_code / targetLanguageCode: Set the language your agent should speak, for example hi-IN or en-IN.
• model: Use bulbul:v3.
• speaker: Use a speaker supported by the selected model. The default is shubh for bulbul:v3.
• speech_sample_rate / sampleRate: Use 22050 for general voice agent audio; use 8000 only when your downstream path requires narrowband telephony audio.
• pace: Start at 1.0, then tune after listening to full agent turns.

from livekit.agents import AgentSession
from livekit.plugins import sarvam

session = AgentSession(
   tts=sarvam.TTS(
      target_language_code="hi-IN",
      model="bulbul:v3",
      speaker="shubh",
      speech_sample_rate=22050,
      pace=1.0,
      output_audio_bitrate="128k",
      output_audio_codec="mp3",
      min_buffer_size=50,
      max_chunk_length=150,
      send_completion_event=True,
   ),
   # ... llm, stt, etc.
)

import { voice } from '@livekit/agents';
import * as sarvam from '@livekit/agents-plugin-sarvam';

const session = new voice.AgentSession({
    tts: new sarvam.TTS({
        targetLanguageCode: "hi-IN",
        model: "bulbul:v3",
        speaker: "shubh",
        pace: 1.0,
        temperature: 0.6,
    }),
    // ... llm, stt, etc.
});

Parameters

This section describes commonly used parameters. See the plugin reference links in the Additional resources section for a complete list of all available parameters.

Speakers

Speaker availability depends on the selected model. The following lists show all speakers supported by the Python plugin. The Node.js plugin supports additional bulbul:v3 speakers not listed here. For the most up-to-date list, see How to change the speaker .

bulbul:v3

The default speaker is shubh.

Female: amelia, ishita, kavitha, kavya, neha, pooja, priya, ritu, roopa, rupali, shruti, shreya, simran, sophia, suhani, tanya.

Male: aayan, aditya, advait, amit, ashutosh, dev, kabir, manan, rahul, ratan, rohan, shubh, sumit, varun.

bulbul:v2

The default speaker is anushka.

Female: anushka, arya, manisha, vidya.

Male: abhilash, hitesh, karun.

Troubleshooting

Common issues and solutions for the Sarvam TTS plugin.

Unsupported speaker or model

If the plugin rejects your configuration, check the model and speaker combination. Speaker availability depends on the selected model, and some parameters are model-specific.

Audio starts too slowly

For streaming voice agents, review chunking and buffering first:

• Reduce min_buffer_size gradually if the agent waits too long before speaking.
• Reduce max_chunk_length if long LLM responses are delaying synthesis.
• Keep punctuation in the generated text so the TTS system can split speech naturally.
• Avoid changing several latency-related settings at once.

Speech sounds rushed, slow, or unnatural

Start with pace=1.0 and temperature=0.6, then tune one setting at a time. If the agent speaks long paragraphs, consider splitting the LLM response into shorter, conversational sentences before it reaches TTS.

Output format does not match your media path

Check speech_sample_rate, output_audio_codec, and output_audio_bitrate. Browser playback, mobile playback, and telephony paths often need different formats. For phone calls, confirm whether your provider expects 8000 Hz audio, mulaw, alaw, or linear PCM.

Pronunciations are inconsistent

For bulbul:v3, use dict_id when you need consistent pronunciations for names, brands, product terms, acronyms, or domain-specific words, provided you have an existing Sarvam TTS pronunciation dictionary.

Additional resources

The following resources provide more information about using Sarvam with LiveKit Agents.

Python

PyPIReferenceGitHub

Node.js

NPMReferenceGitHub

Sarvam docs

Sarvam's full docs site.

Sarvam TTS API reference

Sarvam's text-to-speech API documentation.

Voice AI quickstart

Get started with LiveKit Agents and Sarvam.

Sarvam STT

Guide to the Sarvam STT plugin with LiveKit Agents.