Sesame published its method for a voice that feels present

What happened

On 27 February 2025, the voice start-up Sesame published a research post, Crossing the uncanny valley of conversational voice, introducing its Conversational Speech Model (CSM). The post frames the goal as 'voice presence,' which it defines as 'the magical quality that makes spoken interactions feel real, understood, and valued.' Weeks later the company released a smaller version of the model on GitHub, dated 13 March 2025 in the repository's own history.

What the documents show

The research post describes CSM as a transformer that 'operates directly on RVQ tokens,' processing text and audio together and drawing on conversation history to shape how a reply sounds. It states that an effective voice companion needs 'emotional intelligence, natural conversational dynamics, contextual awareness, and consistent personality,' and commits to 'open-sourcing key components of our research' under an Apache 2.0 license. The GitHub repository confirms that license and states the released model has 1 billion parameters, uses 'a Llama backbone and a smaller audio decoder' to produce audio codes, cannot generate text on its own, lacks voice-specific fine-tuning, and has limited support for languages other than English. Separately, Sesame's own homepage describes a different, unreleased product: smart eyewear intended for 2027, positioned as the eventual home for this kind of voice agent.

The mechanism

'Voice presence' is an experience Sesame is designing toward, not a measurement it reports; the technical substance behind that phrase is a model that conditions each generated audio token on the preceding conversation, rather than treating each reply as a fresh, context-free utterance. That is a concrete architectural claim, distinct from the more evocative language of feeling 'real, understood, and valued,' which describes an intended effect on a listener rather than a property of the code. Readers evaluating a presence claim should ask which half of it, the conditioning mechanism or the described feeling, a given source is actually supporting.

What it leaves open

The research post does not publish latency figures comparable to Moshi's, nor an evaluation methodology for 'voice presence' itself. The GitHub release is explicitly smaller and more limited than whatever model might power a shipped Sesame product on the eyewear the company describes elsewhere, so the openly available artifact and any future commercial voice may differ substantially.

• Does a 'presence' claim describe a measured property of the audio, or a described effect on the listener?
• Is the model being evaluated the same one a company plans to ship, or a smaller research release?
• What conversation history is the model actually conditioning on, and how long does that context persist?

Sesame's contribution is a named design goal and a released method for pursuing it, not yet a finished, measured companion voice product.