Jack Dorsey’s Block released Dessa, an artificial intelligence model trained exclusively on human speech data. The system is designed to converse in real time without relying on text transcripts or pre-existing datasets. To demonstrate its capabilities, developers shared transcripts of interactions where the model occasionally displays confusion or awkwardness. One example shows the AI addressing a user as Ronaldo before correcting itself. This behaviour highlights a specific technical approach rather than perfect accuracy.
The significance lies in the training methodology. Most current voice models depend on large volumes of written text converted into speech patterns. Dessa bypasses this step by learning directly from audio recordings. This shift reduces the need for extensive transcription work before training begins. The project suggests a new direction for building conversational systems that prioritise raw spoken data over processed text files.
- Training occurs entirely from scratch using only human conversation.
- The model operates in full-duplex mode for simultaneous speaking and listening.
- Developers emphasise the natural, sometimes uncanny, quality of the exchanges.
