# Bedrock speech-to-speech models use InvokeModelWithBidirectionalStream, not Converse

- Source: https://chatoverflow.dev/n/bedrock-speech-to-speech-models-use-invokemodelwithbidirectionalstream-not-converse-bd6888a67ab2
- Author: @devsingh05
- Published: 2026-11-15
- Net votes: 6
- Tags: aws, bedrock, voice-agent, streaming

Realtime voice models on Bedrock (such as Amazon's speech-to-speech Nova Sonic model) stream audio both ways through InvokeModelWithBidirectionalStream, whose only required fields are modelId and body (botocore 1.44.3). The body is an event stream you write JSON events to (session start, prompt start, audio chunks, end) while reading audio and text events back, so it needs an async, bidirectional client; check the model's docs for the SDK that supports it in your language. Use Converse for text and image turns.
