Comsync Voice API

Streaming text-to-speech for Hindi, Hinglish and English, built for voice agents. Send text, get natural speech back while it is still being generated.

Write Hindi in Devanagari (आपका order), not in Latin letters (aapka order). The API converts common romanized Hindi automatically, but Devanagari always sounds best. If your LLM writes the reply, tell it to use Devanagari for Hindi.

Authentication

Every request needs an API key, sent as a bearer token. Request a key if you don't have one yet.

Authorization: Bearer $COMSYNC_TTS_API_KEY

Browsers can't set headers on WebSockets, so the streaming endpoint also accepts ?api_key= in the URL. Keep keys on your server; never ship them in a public web page or app.

Quickstart

Generate your first clip. The response streams as it is produced, so you can start playing before the sentence is finished.

curl https://api.trycomsync.com/v1/audio/speech \
  -H "Authorization: Bearer $COMSYNC_TTS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": "नमस्ते! आपका appointment confirm हो गया है।", "voice": "siya_lively_best"}' \
  -o hello.mp3

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.trycomsync.com/v1", api_key=os.environ["COMSYNC_TTS_API_KEY"])

with client.audio.speech.with_streaming_response.create(
    model="comsync-run11",
    voice="siya_lively_best",
    input="नमस्ते! आपका appointment confirm हो गया है।",
) as response:
    response.stream_to_file("hello.mp3")

import fs from "node:fs";

const res = await fetch("https://api.trycomsync.com/v1/audio/speech", {
  method: "POST",
  headers: {
    "Authorization": `Bearer ${process.env.COMSYNC_TTS_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({ input: "नमस्ते! आपका appointment confirm हो गया है।", voice: "siya_lively_best" }),
});
await fs.promises.writeFile("hello.mp3", Buffer.from(await res.arrayBuffer()));

LiveKit plugin

The plugin streams your LLM's text into Comsync Voice as it is written and plays audio from the first phrase. When the caller interrupts, it closes the stream and generation stops on our side within one audio frame.

1. Download livekit_comsync.py and put it next to your agent. It needs livekit-agents 1.2 or later.
2. Set COMSYNC_TTS_API_KEY in your agent's environment.
3. Pass it as the session's TTS:

from livekit.agents import AgentSession
from livekit_comsync import TTS

session = AgentSession(
    stt=...,
    llm=...,
    tts=TTS(voice="siya_lively_best"),  # reads COMSYNC_TTS_API_KEY
)

OpenAI-compatible clients

POST /v1/audio/speech follows OpenAI's speech API, so OpenAI SDKs and LiveKit's OpenAI plugin work by changing the base URL. The plugin speaks each sentence once the LLM has finished it, rather than streaming text in, so first audio is a little later than with the Comsync plugin.

import os
from livekit.plugins import openai

tts = openai.TTS(
    base_url="https://api.trycomsync.com/v1",
    api_key=os.environ["COMSYNC_TTS_API_KEY"],
    voice="siya_lively_best",
    response_format="pcm",
)

Create speech

Generates speech from text and streams it back as it is produced.

Responses include X-Request-Id, X-Sample-Rate and X-Channels headers. Quote the request id if you contact support about a clip.

Streaming WebSocket

For voice agents: send text as your LLM produces it and receive audio as soon as each phrase is ready. One connection can serve a whole call, turn after turn.

Messages you send

Messages you receive

Voices

Each voice is set by a short reference recording and tuned for conversational calls. Play any of them below; GET /v1/voices returns the full list of 24 ids with names and descriptions.

Errors

Errors use OpenAI's shape: {"error": {"message", "type", "code"}}. The message always says what to change.

WebSocket connections are refused with close code 4401 for a bad key and 4429 when the key is at its limit.

Limits and performance

• First audio: about 200 ms on the server, about 300 ms end to end from Indian networks.
• Generation speed: about 2.4× faster than real time per stream, so playback never waits mid-sentence.
• Concurrency: each key has a limit on simultaneous requests, agreed when you get access. Ask us before running load tests.
• Input: up to 4,000 characters per HTTP request. For conversations, send one reply at a time over the WebSocket.
• Accuracy: 1.4% character error rate on our Hindi test set, measured with an independent speech recogniser.

Support

Questions, higher limits or a dedicated deployment: email ojas.jain@comsync.in with your request id if you have one.