Trained only on audio, without text
Payments company Block released "Dessa," a conversational AI that listens and responds in voice, as a research version on Oct. 23 (Oct. 24 Japan time). Jack Dorsey, the company's co-founder and chairman, introduced it on X.
Many voice AIs are built on language models trained on text, converting speech to text to produce answers. Dessa was reportedly trained only on audio of human conversation, without text data or pretrained models. The company says it uses a "full-duplex" design that lets it speak while listening to the other person, enabling conversations that feel like talking with a person, including emotion, pauses and intonation.
Company claims it needs fewer words
Block says the number of words needed for training is a few hundredths of what a language model requires. However, when the same words are recorded as audio, the amount of data is more than 1,000 times that of text, it says. The company has not disclosed the amount of data used for training or the number of parameters.
Developed by a small team from Toronto
It was developed by a small research lab inside Block. The lab was founded by people who had worked at "Dessa," a deep learning startup in Toronto, Canada, and is led by Ragavan Thurairatnam. It consists of six researchers plus interns and others, and reports directly to Dorsey.
No evaluation figures, and limited access
Block has not provided an API, pricing or an official release date. For the general public, it is only accepting waitlist sign-ups, and a small number of AI safety researchers will test it first to look for unanticipated risks. Its performance is the company's claim, and there are no evaluation figures or third-party verification yet. One of the published samples is a conversation with OpenAI's voice AI "GPT-Live-1," but Block noted that it has no affiliation with OpenAI.
