Amazon Releases Nova 2.5 Sonic Speech-to-Speech Model for Real-Time Voice Agents
The model supports seven languages and is generally available on Amazon Bedrock at the same price as its predecessor.
Amazon Web Services has released Nova 2.5 Sonic, a speech-to-speech model built for real-time voice agents such as customer service lines and in-app assistants.
Key details
- Supports seven languages
- Generally available on Amazon Bedrock in U.S., European and Asian regions
- Priced the same as Nova 2 Sonic
Speech-to-speech models skip the traditional pipeline of transcribing speech, generating a text reply and converting it back to audio. Handling voice end to end typically cuts latency and keeps more of the speaker's tone, which matters for natural back-and-forth conversation.
Agent framework goes GA
AWS also made its Strands Bidi Agents framework generally available. It is designed for building agents that handle bidirectional, streaming conversations, a natural pairing with the new voice model.
The launch is part of a broader AWS push to keep enterprise AI workloads on Bedrock as competition with Microsoft Azure and Google Cloud intensifies ahead of third-quarter earnings.