Voice AI

Real-time Voice AI

Put an AI agent into a live call as a participant

SRTC carries low-latency audio between users and your agent; the agent joins through the server-side Python SDK. Plug in a speech-to-speech model or build your own ASR / LLM / TTS pipeline — for customer service, AI interviews, tutoring or meeting assistants, self-hosted if you need it.

Talk to an expert Talk to the AI agent

Challenges

Human support is expensive and cannot cover every hour

Hosted AI voice services send your data somewhere you do not control

Locked into one AI vendor when a better or cheaper model ships

Handling WebRTC, codecs, echo and barge-in yourself is a lot of work

Phone systems and AI live in separate silos

What we deliver

Speech-to-speech models

Connect Gemini Live, OpenAI Realtime, Qwen Omni or Doubao realtime speech for replies in about a second and the most natural, interruptible turns.

Cascaded ASR → LLM → TTS

Mix Deepgram or FunASR, any LLM, and CosyVoice or ElevenLabs — swap each stage on quality and cost.

Python SDK + pipecat

pip install srtc, join a channel, read and write PCM. A built-in pipecat transport drops into existing pipelines.

Barge-in

TTS audio is paced out in 20 ms frames; clear the buffer instantly when the user interrupts.

Multi-party

The agent is a regular participant: it hears every speaker separately, ideal for meeting assistants and panel interviews.

Phone lines, too

Through SIP on the Room Connector, callers dial straight into your agent — AI phone support and outbound calling on the same stack.

Conversation analytics

Recordings, transcripts, sentiment analysis and service-quality scores.

Self-hostable

Run both the RTC layer and your AI stack on-prem; with self-hosted models, no data leaves your network.

A few lines to put your agent on the call

Server-side Python SDK for Linux x86_64 / aarch64, macOS and Windows, Python 3.10+.

# pip install "srtc[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()])

Python SDK docs

Why SRTC

Replies in about a second, interrupt anytime

Speech-to-speech or cascaded — your choice

SIP phone calls straight into the agent

No AI vendor lock-in

Python SDK with pipecat transport

Barge-in and multi-party ready

Fully self-hostable

Clients use SRTC SDKs on every platform

FAQ

Which AI models are supported?

SRTC is model-agnostic. For speech-to-speech, connect Gemini Live, OpenAI Realtime, Qwen Omni or Doubao realtime speech. For cascaded pipelines, combine any ASR (Deepgram, FunASR…), LLM and TTS (CosyVoice, ElevenLabs…) — most services supported by pipecat work out of the box.

How does my agent connect to SRTC?

Use the server-side Python SDK: pip install srtc (or srtc[pipecat]). Your backend issues a token through the server API, the agent joins with it, receives each user’s decoded PCM frames and writes TTS audio back to the channel.

Speech-to-speech or cascaded?

Speech-to-speech gives the lowest latency and most natural voice. Cascaded pipelines let you replace each stage with a self-hosted model for compliance or cost. Both run on the same SRTC deployment.

Can the agent take phone calls?

Yes. With SIP access on the Room Connector, phone lines can dial straight into your agent and the agent can place outbound calls — for AI phone support, follow-up calls and similar use cases.

Ready to build?

Get a self-hosted trial, SDK access and hands-on help from our engineers