Build a realtime voice agent

Experimental Feature

This feature is experimental and may change in future versions. Use with caution in production environments.

A voice agent listens and talks at the same time: audio streams in while the model streams a reply out, the user can cut in mid-sentence, and tools run without pausing the conversation. Bidirectional streaming is the piece that makes this work. Instead of the request-then-response cycle of a standard agent, a BidiAgent holds a persistent connection to a realtime model and moves audio, text, and events across it continuously. This section covers building that agent, choosing a model provider that supports it, and handling the live stream.

Get started

Build a voice agentStand up a listening, talking agent end to end, from install to first conversation.

BidiAgentConfigure the persistent-connection agent: model, tools, system prompt, and lifecycle.

Model providers

Nova SonicAmazon's bidirectional streaming model, served through Amazon Bedrock.

OpenAI RealtimeOpenAI's Realtime API, with a longer connection timeout for extended sessions.

Gemini LiveGoogle's multimodal Live API for streaming voice and video.

Handle the live stream

I/O channelsWire local audio and text I/O, or implement custom channels for a web server.

Streaming eventsHandle the events the agent emits as audio, transcripts, and tool calls arrive.

InterruptionsLet a user cut in mid-response and have the agent stop and listen.

A streaming voice agent

The smallest real thing this section builds: a BidiAgent on a realtime model, reading from the microphone and playing back through the speakers. The run() loop streams audio both ways until you interrupt it.

import asyncio

from strands.experimental.bidi import BidiAgent, BidiAudioIO
from strands.experimental.bidi.models import BedrockNovaSonicModel

model = BedrockNovaSonicModel()
agent = BidiAgent(
    model=model,
    system_prompt="You are a helpful voice assistant. Keep replies short and natural.",
)
audio_io = BidiAudioIO()


async def main():
    # Stream microphone audio in and speaker audio out until interrupted
    await agent.run(inputs=[audio_io.input()], outputs=[audio_io.output()])


asyncio.run(main())

Bidirectional streaming is a Python-only experimental feature; install it with pip install "strands-agents[bidi-all]". The quickstart covers per-provider installs and credentials.

Where to go next

New here? Start with the quickstart to get a voice conversation running, then read BidiAgent to configure tools, prompts, and the connection lifecycle. Pick a model provider based on the provider you use and the session length you need.

Building for a server rather than a local machine? Read I/O channels to replace microphone-and-speaker I/O with your own transport, then streaming events and interruptions to drive the conversation from your own event loop.

Tags

bidi-streaming11