AWS published a SageMaker AI implementation that synthesizes agent responses sentence by sentence, reporting faster first-audio delivery than batch synthesis.
On October 22, 2026, AWS published a technical implementation for streaming spoken responses from AI agents using Kokoro-82M on Amazon SageMaker AI. Instead of waiting for an agent to finish generating a full response before starting text-to-speech, the design sends completed sentences to the speech model as they become available.
The 82-million-parameter model runs in a custom container on a SageMaker AI real-time endpoint using an ml.g5.xlarge GPU instance. The example connects a Strands agent, backed by Amazon Bedrock and deployed in Amazon Bedrock AgentCore, to that endpoint. AWS describes two endpoint routes: one for streaming a response to a single input and another bidirectional route that keeps a session open across multiple messages.
In five trials using the same 267-character, four-sentence text, AWS measured mean time to first audio byte of 156.7 milliseconds when sending the first sentence incrementally, versus 329.3 milliseconds when sending the whole text as one batch. In a separate end-to-end test, the bidirectional approach produced first audio 2.4 times faster than batch mode in a local container and 3.5 times faster on the deployed AgentCore runtime. These are results from AWS’s described test setup, not a general benchmark across models or deployments.
The article also compares ways to deliver audio to a browser. Base64-encoded audio in Server-Sent Events can reuse an existing text-streaming pipeline, but adds about 33% to payload size and requires buffering chunks. A bidirectional connection can instead carry raw audio alongside JSON control messages.
AWS flags a security caveat in its sample: it pins a PyTorch version with a known advisory, fixed in torch 2.10.0, and advises upgrading that dependency before use beyond proof-of-value testing.