How Did Dessa Learn to Talk? Training a Voice AI Without Text
Block says Dessa learned to hold a conversation from human recordings alone, with no text and no pretrained model. What does that change compared with the voice AIs you already talk to?
- Dessa’s distinctive announcement is Block’s claim that it learned from human conversation audio alone.
- Full-duplex speech models already exist, including designs that use text.
- Conversational timing, task execution and product access answer different questions.
The interesting claim concerns how the model learned
Block describes Dessa as trained from scratch on human conversation recordings, without text or pretrained models. That is a claim about its learning process, beyond the familiar idea of speaking to an AI and hearing a spoken answer.
Two questions follow. Can the system listen while it speaks? And what training produced its language abilities? A convincing conversation can demonstrate behavior without revealing the answer to the second question.
Text-based foundations can also support simultaneous listening and speaking
Full duplex describes overlapping input and output, not an audio-only training requirement. Kyutai’s 2024 Moshi paper starts with a text language model and models the user’s speech and the assistant’s speech in parallel. It supports interruptions and overlapping conversation without fixed speaker turns.
The official architecture description also includes text tokens corresponding to the model’s own speech. Text can therefore participate inside a system that presents a direct speech-to-speech experience.
| Term | What it describes | What it does not establish |
|---|---|---|
| Speech-to-speech | Spoken input and spoken output. | Whether text was used in training. |
| Full duplex | Listening and speaking simultaneously. | The accuracy of every answer. |
| Speech-only training | Learning from audio alone. | Support for documents or external tools. |
The useful comparison is consequently what audio-only training enables and where its knowledge reaches. The existence of full-duplex behavior alone does not establish a new capability or superiority over another model.
An expressive conversation is not a completed transaction
A service assistant that checks an order or changes a booking needs an interface to external systems. Natural interruptions and expressive delivery do not, by themselves, validate the arguments sent to a booking API.
The NemotronLabs VoiceChat paper illustrates that engineering distinction. Its architecture includes structured function-call output alongside conversational streams, allowing tool use during an ongoing spoken exchange. The authors still identify argument accuracy and end-to-end tool execution as areas for improvement.
For a customer-service assistant, evaluate conversational timing, factual answers and completed external actions separately. Keeping a conversation flowing and completing an accurate order lookup are different selection criteria.
A Mac capability claim is different from an installation path
Block says Dessa runs locally on an M4 Max MacBook Pro or better. As of October 26, its access information points to a waitlist and a small group of AI safety researchers. The page does not specify memory requirements or a Korean-language support commitment.
Developers seeking a runnable full-duplex research system can inspect Moshi’s published models and PyTorch, MLX and Rust implementations. Those artifacts provide a concrete starting point for checking hardware requirements and running experiments.
Dessa is an interesting test of what conversation audio alone can teach. Choosing it for a product requires a further match between language support, the task to be completed and the access actually available. Those are more useful decision points than how human a voice seems.