Amazon's New AI Voice Tech Cuts Awkward Silences in Half

The article describes a way to close the gap in AI voice agents: instead of waiting for a full multi-paragraph response before text-to-speech starts, it streams each sentence to a hosted Kokoro-82M TTS model on Amazon SageMaker AI as soon as the agent produces it, so synthesis can begin on sentence one and audio playback can start before the full response is finished. The article says a one-sentence response might generate in 600 milliseconds, but the full response can take five seconds, and a typical integration waits for all of it before synthesis starts.

Kokoro-82M has only 82 million parameters and runs in a custom container on a SageMaker AI real-time endpoint on an ml.g5.xlarge GPU instance, with one route for streaming a response for a complete input and one bidirectional WebSocket route that keeps a connection open across many messages. The walkthrough deploys the endpoint and an Amazon Bedrock AgentCore runtime hosting a Strands agent, then covers client-side changes, including handling binary WebSocket frames, keeping the session open between messages, and streaming sentences as the agent produces them.