Block unveils speech-trained AI model for more natural conversations
Block has introduced Dessa, a voice AI model trained from scratch on human conversation, and says it can speak directly with people without converting their words into text. The project explores whether learning from speech alone can make AI conversations feel more natural and help systems recognise emotion, which Block argues could improve safety.
Demonstrations show mostly fluent exchanges, but also awkward moments: Dessa initially mistakes a speaker for someone named Ronaldo and veers off topic during a discussion of films. Block says the model can run locally on a MacBook Pro M4 Max or better, and plans to give early access to a small group of AI safety researchers.
In short
- Block unveiled its speech-to-speech model Dessa.
- It learns from human speech without written text.
- Early access will go to AI safety researchers.
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Training AI directly on speech could make conversations feel more fluid by preserving tone, emotion and other cues that text transcription may lose. Local operation and early access for safety researchers could also support privacy and scrutiny as the technology develops.
The case against
Speech-only training does not by itself ensure that a system understands people reliably or responds safely; the reported identity mix-up and off-topic reply illustrate current limits. A new approach may be premature to promote as safer or more natural until it is tested more broadly, including how it handles errors and sensitive conversations.
Explained simply
Block is a financial technology company whose products include payment services and the Cash App. Its Dessa project explores a different way to build conversational AI: the model learns from speech and responds directly with speech, rather than first turning spoken words into text.
Most voice assistants rely on text-based steps to process what people say. Block says learning from speech alone may help a system pick up emotion and make conversation feel more natural, while also raising questions about how to make voice AI safe and reliable. Early access is intended for a small group of AI safety researchers.
Read the full article at the source →
Originally published by The Verge as “Jack Dorsey’s Block made an AI model that talks to you like an awkward new friend”.