Amazon Nova Sonic

What it is

Amazon Nova Sonic is a speech-to-speech foundation model in the Nova family, announced on 8 April 2025 and made available through Amazon Bedrock^[1]. Its distinguishing feature is that it does not treat a spoken conversation as a series of separate jobs. Speech recognition, language understanding and speech generation all happen inside one model, rather than being handed off between a transcription model, a language model and a text-to-speech system^[1].

That design choice is the whole argument for the model. AWS described the conventional arrangement — chaining a speech recognition component, a language model and a text-to-speech component — as fragmented, and said such pipelines tend to lose tone, prosody and speaking style along the way^[1]. A single model that hears audio and produces audio keeps those qualities inside the same system. AWS said the unified architecture delivered expressive speech generation and real-time text transcription without needing a separate model, at what it described as industry-leading price performance^[1].

How developers reach it

Nova Sonic is used through a new bidirectional streaming API, InvokeModelWithBidirectionalStream, running over HTTP/2, under the model ID amazon.nova-sonic-v1:0^[1]. Streaming in both directions is what makes a natural back-and-forth possible: audio goes in while audio comes out, rather than the user speaking, waiting for a complete transcript, and waiting again for a synthesised reply.

The model has a context window of 300K tokens, and the default connection limit is 8 minutes^[1]. It is built to handle interruptions and background noise without losing the thread of the conversation^[1] — the small, ordinary things that make a voice interface feel usable rather than brittle.

Beyond conversation, Nova Sonic supports real-time transcription and tool use, including retrieval-augmented generation with Bedrock Knowledge Bases^[1]. That means a spoken application built on it can look things up in a connected knowledge source mid-conversation instead of answering only from what the model already holds.

Languages and safeguards

At launch the model supported American and British English^[1]. Spanish was added on 12 June 2025, and French, Italian and German followed on 16 July 2025^[1]. The July update also brought six expressive voices with masculine- and feminine-sounding options^[1].

The model includes built-in content moderation and watermarking, and it launched in the US East (N. Virginia) region^[1]. Watermarking matters for a system whose output is synthetic speech: it gives downstream systems and listeners a way to identify generated audio.

Where it sits

Nova Sonic represents a step in how voice interfaces are assembled — from a pipeline of specialised models glued together, to one model that carries a conversation in audio end to end. The practical claims attached to it are about latency, expressiveness and cost rather than about new capabilities in reasoning.

Its place in the family was short-lived at the top. Nova Sonic was succeeded by Nova 2 Sonic, announced in December 2025^[2]. The first model is best understood as the point where Amazon moved speech-to-speech into a single foundation model and opened it to developers through Bedrock, with later work building on that base.