Nova Sonic

Nova Sonic is Amazon's speech-to-speech foundation model, announced in April 2025 on Amazon Bedrock. Instead of chaining separate speech-to-text, language, and text-to-speech systems, it unifies speech understanding and generation in a single model, preserving acoustic context like tone, pace, and style that pipeline approaches lose. It runs over a bidirectional streaming API with a 300,000-token context window.

Amazon positions it for real-time voice agents in customer service and call centers, with tool use and knowledge grounding built in so agents can call functions mid-conversation. Reported benchmarks include a 4.2 percent word error rate on Multilingual LibriSpeech and win rates against Gemini Flash 2.0 and GPT-4o in single-turn English conversations, though those are vendor-reported. Spanish support followed in June 2025, and a Nova 2 Sonic successor later expanded languages and context. Access is API-only through Bedrock.