SRTC Python SDK

English | 中文

SRTC is a real-time audio/video engine built for fully private deployment, with SDKs for every major platform, support for domestic (Xinchuang) OS and hardware, interoperability with legacy SIP/H.323 devices, and built-in AI Agent capabilities.

This SDK is for server-side AI: voice agents (ASR → LLM → TTS), recording, transcription and QA. Full documentation: docs.stmlink.com.

• PCM in, PCM out — remote audio arrives decoded at the sample rate you choose; write TTS output as raw PCM. Codecs and resampling are handled inside the SDK.
• Real-time pacing built in — however fast your TTS is, audio goes out one 20 ms frame at a time; clear() stops playback instantly when the user barges in.
• Continuous timeline — silence the remote side suppresses (DTX) is filled back in by RTP timestamp, so VAD and recording durations stay correct.
• asyncio native — async with to leave automatically, async for to consume frames, sync or async event handlers.
• Native core — signaling, reconnection and media transport run in a bundled native library, behaving the same as every other SRTC SDK and staying off the GIL.
• pipecat ready — a built-in pipecat transport drops into existing voice-agent pipelines.

Install

pip install srtc                # core
pip install "srtc[pipecat]"     # with pipecat integration

Python 3.10+. Linux x86_64 / aarch64, macOS Apple Silicon, Windows x64. Dependencies: cffi, av (bundles ffmpeg and libopus — no system ffmpeg needed), numpy.

Linux with glibc < 2.28 (e.g. CentOS 7): the latest av only ships wheels for glibc 2.28+. Pin it: pip install srtc "av>=12,<14".

Quick start

import asyncio
import srtc

async def main(token: str):   # token is issued by your backend via the SRTC server API
    async with await srtc.Channel.join(
        token,
        auto_subscribe_audio=True,                  # hear everyone, including later joiners
        audio_format=srtc.AudioFormat(16000, 1),    # PCM format you receive (default 16 kHz mono)
    ) as ch:
        tts = await ch.publish_audio(desc="tts", audio_format=srtc.AudioFormat(24000, 1))

        async for frame in ch.audio_frames():       # frame.uid tells you who is speaking
            text = asr.feed(frame.uid, frame.to_numpy())
            if text:
                await tts.write(await tts_engine.synthesize(await llm.chat(text)))

asyncio.run(main(token))

Each Channel.join needs a freshly issued token — a token is bound to one session.

Sending audio

await tts.write(pcm)            # any length; paced out in real time, back-pressure when the buffer is full
tts.clear()                     # user barged in: drop everything not yet sent
await tts.wait_for_playout()    # wait until everything written has been sent
tts.buffered_seconds            # > 0 while the agent is still "speaking"

Events — subclass srtc.ChannelHandler and override what you need (on_user_join, on_user_leave, on_track_added, on_audio_frame, on_active_speakers, on_custom_msg, on_disconnected, …), then pass handler= to Channel.join.

pipecat

from srtc.pipecat_transport import SRTCParams, SRTCTransport

transport = SRTCTransport(token, SRTCParams(audio_in_enabled=True, audio_out_enabled=True))
pipeline = Pipeline([transport.input(), stt, llm, tts, transport.output()])

Errors — failures raise srtc.SdkError with code and msg: 180xxx are SDK errors, codes ≥ 1000 come from the server (e.g. 1033 concurrency limit reached).

Deployment notes — don't fork a process that has already created a Channel (use spawn with multiprocessing; importing before fork is fine). A single process handles about 20 concurrent channels at ~61% of one core; scale out with more processes.

See the documentation for the full API, the voice-agent guide and the pipecat guide.

License

Proprietary. Copyright © Seastart. Use requires an SRTC service license; copying, modification or redistribution without permission is prohibited. A free edition with limited concurrency is available; higher concurrency and private deployments require a commercial license — see stmlink.com.

---

SRTC Python SDK（中文）

English | 中文

SRTC 是一款支持完全私有化部署的实时音视频通信引擎，覆盖全平台 SDK，支持信创国产化，兼容传统 SIP/H.323 设备，集成 AI Agent 能力。

本 SDK 面向服务端 AI 场景：语音 agent（ASR → LLM → TTS）、录音、转写、质检。完整文档见 SRTC 文档中心。

• 收发都是 PCM：远端音频按你指定的采样率解码后交给你；TTS 输出的 PCM 直接写进去，编解码、重采样都在 SDK 内完成。
• SDK 控发送节奏：TTS 生成再快，也按实时节奏每 20ms 发一帧；用户插话时 clear() 立即停止播报。
• 时间轴连续：对端 DTX 静音不发包时，SDK 按 RTP 时间戳补回静音帧，VAD 断句和录音时长都不会错位。
• asyncio 原生：async with 自动离开、async for 逐帧消费，事件回调可以是普通函数也可以是 async 函数。
• 原生内核：信令、断线重连、音视频收发由内置原生库完成，与 SRTC 其它端行为一致，不占 Python GIL。
• 接入 pipecat：内置 pipecat transport，直接接进现有的语音 agent 管线。

安装

pip install srtc                # 核心
pip install "srtc[pipecat]"     # 同时安装 pipecat 集成

Python 3.10+。平台：Linux x86_64 / aarch64、macOS Apple Silicon、Windows x64。依赖：cffi、av（自带 ffmpeg 与 libopus，无需系统安装 ffmpeg）、numpy。

glibc 低于 2.28 的 Linux（如 CentOS 7）：av 的新版本只提供 glibc 2.28+ 的安装包，请固定版本安装：pip install srtc "av>=12,<14"。

快速开始

import asyncio
import srtc

async def main(token: str):   # token 由你的业务服务端通过 SRTC 服务端 API 签发
    async with await srtc.Channel.join(
        token,
        auto_subscribe_audio=True,                  # 自动订阅所有人（含之后加入者）的音频
        audio_format=srtc.AudioFormat(16000, 1),    # 收到的 PCM 格式，默认 16k 单声道
    ) as ch:
        tts = await ch.publish_audio(desc="tts", audio_format=srtc.AudioFormat(24000, 1))

        async for frame in ch.audio_frames():       # 用 frame.uid 区分说话人
            text = asr.feed(frame.uid, frame.to_numpy())
            if text:
                await tts.write(await tts_engine.synthesize(await llm.chat(text)))

asyncio.run(main(token))

每次 Channel.join 都要用新签发的 Token，一个 Token 只绑定一次会话。

发送音频

await tts.write(pcm)            # 任意长度，按实时节奏发送；缓冲满时 write 会等待（背压）
tts.clear()                     # 用户插话：丢弃所有未发出的音频
await tts.wait_for_playout()    # 等待已写入的全部发完
tts.buffered_seconds            # 大于 0 表示 agent 还在"说话"

事件回调：继承 srtc.ChannelHandler，覆写需要的方法（on_user_join、on_user_leave、on_track_added、on_audio_frame、on_active_speakers、on_custom_msg、on_disconnected 等），在 Channel.join 时通过 handler= 传入。

pipecat

from srtc.pipecat_transport import SRTCParams, SRTCTransport

transport = SRTCTransport(token, SRTCParams(audio_in_enabled=True, audio_out_enabled=True))
pipeline = Pipeline([transport.input(), stt, llm, tts, transport.output()])

错误：失败时抛 srtc.SdkError，带 code 与 msg。180xxx 为 SDK 自身错误，≥1000 为服务端错误码（如 1033 并发已达上限）。

部署注意：已创建过 Channel 的进程不要再 fork（multiprocessing 请用 spawn；fork 前只 import 没问题）。单进程约 20 个频道同时在线时占约 61% 单核 CPU，会话更多请多进程横向扩展。

完整接口、语音 agent 实践与 pipecat 接入见 文档中心。

许可证

专有软件，版权归 Seastart 所有。使用需获得 SRTC 服务授权，未经许可不得复制、修改或再分发。SRTC 提供免费版（并发量受限），超出限额的并发与私有化部署需购买商业授权，详见 官网。