A Voice Model That Never Read a Word
Block trained Dessa on conversation audio alone, with no text and no pretrained model, and made it talk over you. The interesting claim is not that it sounds human. It is that the transcript stopped being the intermediate layer.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page
You are describing a Christopher Nolan film. It is set in space. You are still reaching for the title when the other person says Interstellar. That is a normal conversation between two film fans, except one of them is a model Block demonstrated on 23 October, and the thing that gives it away is not the voice. It is that it did not wait its turn.
The model is called Dessa. Block's claim is specific and unusual: it was trained from scratch on human conversation, on audio only, with no text and no pretrained language model underneath. The demos are the kind of thing that makes people on social media insist they are hearing a person. They are also, as Numerama noted, clips Block selected, presented as live-recorded conversations with subtitles added afterwards.
Both things are true. The demos are impressive. They are not evidence.
Full duplex is the actual claim
Most voice assistants run in a relay. Speech recognition decides you have stopped talking, the text goes to a model, the answer comes back as text, and a vocoder performs it. Every stage adds delay, and the handoff at the start is the one you hear as stiffness: the assistant waits for silence that a human would have talked through.
Dessa runs full duplex, meaning it listens and speaks at the same time. It can drop a yes while you are still explaining, laugh before your sentence lands, or step in when you are hunting for a word. Numerama's write-up includes a demo where someone announces a surprise trip to Paris and the model reacts before the announcement is finished.
Full duplex is not new as an idea. Kyutai's Moshi did simultaneous speech in 2024. What Block is claiming is different: Moshi's conversational ability sat on a language model trained on text, while Dessa's training data was speech from the start. That is the part worth arguing about, because it changes what the model can know.
What a transcript throws away
Say ça va brightly and say it with a shaking voice. Identical words, opposite information. A pipeline that converts speech to text before reasoning has to throw that difference away at the door, then hope the vocoder can put it back. Usually it cannot, because the text it is performing carries no record of how the original was said.
Block's stated goal is to keep the prosody in the model rather than reattaching it later: intonation, pauses, sighs, laughter, learned alongside language instead of after it. The company also frames a safety use for this, arguing that an assistant should be able to react to distress audible in a voice even when the words do not say it clearly. That is a reasonable ambition and, at the moment, an untested one. Reproducing an emotional voice in a demo does not show that a model can recognise a state and respond appropriately to it.
For anyone building characters rather than assistants, the practical question is narrower. If a model never had a text layer, what happens to the layer you edit?
The trade creators will actually feel
Every voice pipeline I have worked with keeps text in the middle, and it is there for reasons that have nothing to do with acoustics:
- The script is the version history. You can diff take four against take three and see the line that changed.
- The script is the translation. Dubbing, localisation and subtitles all read from it.
- The script is the review artefact. A director, a client or a safety reviewer can read it without listening to twenty minutes of audio.
- The script is the retake. Fixing one sentence costs one generation, not a re-recording of the scene.
A speech-only model collapses that layer. You get a performance directly from a conversational context, which is exactly why it feels alive and exactly why it is harder to control. The voice design workflow we covered earlier runs the other way: write the voice as text first, cast it, then fix it in post. That approach is slower to sound human and much faster to revise.
The reasonable position, if a model like this reaches production, is to keep both. Generate the performance from audio-native context, then transcribe the result and treat the transcript as the edit surface. You lose the guarantee that the text produced the sound, but you keep the ability to find, compare and license what you shipped.
What a demo cannot tell you
Block is presenting Dessa as a research preview with no launch date and no price, with early access going to a small group of AI safety researchers tasked with finding risks the team did not anticipate. The company says the model runs locally, without the cloud, on a MacBook Pro with an M4 Max or better.
That last claim is the one I would test first, because on-device conversation is a different engineering problem from a good demo. A laptop with a top-tier Apple chip is not a phone, and background noise is not a quiet room.
Four things would separate a demo from a capability:
- Held-out conversations instead of selected clips, including the ones where the model talks over someone at the wrong moment.
- Interruption behaviour under noise, with two people in the room and a television on.
- A published rate for the distress-detection claim, including how often it fires on ordinary tiredness.
- Latency numbers on the hardware it is claimed to run on, measured end to end rather than per component.
None of this is a criticism of the work. It is the normal distance between a compelling recording and a model you can put in a pipeline.
Why it matters for character work
The direction of travel is clear even if this specific model stays a preview. Voice models are moving from "perform this sentence" to "hold this conversation", and that shift rewards a different set of skills. Writing for a model that can interrupt is closer to writing for an actor than writing for a synthesiser: you need beats, not punctuation. Timing becomes part of the script rather than a setting applied afterwards.
It also raises the bar for what counts as a character. A face that holds across images is close enough to solved that we wrote a whole system for it. A voice that holds across an improvised exchange is harder, and a voice that also decides when to speak is harder again.
The part worth keeping is smaller than the headlines. One team stopped treating speech as text with sound attached, and the gap that opened up is where most of a conversation actually lives.