Dauys

The problem

Today's strong full-duplex speech models are a large text language model plus tens to hundreds of thousands of hours of speech, much of it synthesised from text. Text carries the language; speech only teaches the timing. That makes them expensive to train, hard to adapt to languages with little text and synthetic speech — such as Kazakh — and blind to much of what is carried by the voice itself.

The few text-free models that exist learn the rhythm of conversation — who speaks when — but not its content. No open result yet shows that a model can take language and meaning from speech alone at the scale of thousands to tens of thousands of hours.

Our approach

Text helps speech models not because it is text, but because the text stream has three properties: it is slow (3–4 units per second), it carries no voice, and it has low entropy. Dauys v1.0 builds a stream with the same properties directly from audio — syllable-rate units with the speaker's voice removed — and learns conversation on that stream.

The model is split into blocks with hard boundaries enforced by wiring, not by loss penalties. A frozen audio codec turns speech into tokens; a unit encoder compresses them into voice-free syllable units; a causal dialogue tower sees only units and activity of both speakers and is trained mainly to predict the partner; a separate acoustic decoder, which never sees the dialogue, turns the model's own units back into speech in the target voice.

Abilities that are usually mixed together are measured separately: recognising words versus understanding a turn, choosing the right response versus saying it clearly, fluent speech versus actually having heard the other person, and holding a conversation versus knowing about the world.