Amazon introduces Nova Sonic speech-to-speech model

Background

Voice applications have usually been built from several separate models, one for speech recognition, one for understanding the request and one for speech generation. Amazon presented Nova Sonic as a way to combine these steps in a single model for voice conversations in generative AI applications.^[1]

What happened

On 8 April 2025 AWS introduced Amazon Nova Sonic in Amazon Bedrock.^[1] The model combines speech understanding and speech generation, and it streams audio in both directions through a new bidirectional streaming API.^[1]

Key details

The model launched with American and British English voices. Spanish, French, Italian and German were added by July 2025.^[1] Spanish came on 12 June 2025, and the other three on 16 July 2025.^[1] It supports tool use and retrieval-augmented generation, so an application can let the model call external functions and look up material before it answers.^[1] Its context window is 300K tokens.^[1]

A conventional voice system passes audio through separate stages. A speech recogniser turns the sound into text, a language model reads that text and writes a reply, and a speech generator turns the reply back into sound. Each hand-off can lose information such as tone, and it adds delay. A single speech-to-speech model takes audio in and produces audio out, so the exchange can flow more naturally. Interruptions are a good test of this design. When a listener starts talking over the model, the model has to stop, listen and resume with the earlier context still in mind.

For developers, the practical difference is that an application can keep a continuous audio stream open and receive audio back over the same connection, rather than sending a separate request for each step.^[1]

Impact

The launch added a speech-to-speech model to Amazon's Nova family and to Amazon Bedrock. The model was later succeeded by Nova 2 Sonic, which Amazon announced in December 2025.^[1] The model's ID is amazon.nova-sonic-v1:0.^[1]