Amazon Nova Sonic

Amazon Nova Sonic is a speech-to-speech model that Amazon announced in April 2025, unifying speech understanding and speech generation into a single model rather than chaining separate speech-to-text, language, and text-to-speech systems together. It’s accessed through a bidirectional streaming API on Amazon Bedrock, supports a 300,000-token context window, and is built for real-time voice agents that need to listen, reason, and respond with natural-sounding speech and low latency. It also supports tool use and knowledge grounding, so a voice agent built on Sonic can call external functions or retrieve information mid-conversation.

Amazon pitched Sonic as a lower-cost alternative to combining separate ASR, LLM, and TTS pipelines, and as a competitor to OpenAI’s Realtime API and Google’s native audio models for voice assistants and customer service bots. Amazon later followed it with Nova 2 Sonic, which added support for more languages and an expanded context window, but the original Sonic model remains available on Bedrock for existing voice-agent workloads.