Block Unveils Dessa, a Speech-to-Speech AI Model Trained on Conversation
Block unveiled Dessa, an AI model trained only on human conversation and designed for local speech-to-speech use.
The Verge reported that Dessa lead Ragavan Thurairatnam shared one conversation in which the model mostly carried on a natural back-and-forth but initially mistook the speaker for someone named “Ronaldo.” Block’s Dessa website also posted other examples, including a discussion about Christopher Nolan’s films in which Dessa goes on an unexpected tangent about Bryan Singer’s Superman films, and a person telling Dessa about a surprise trip to Paris, to which Dessa “then answers with what sounds like teary gratitude,” as the website describes. A separate conversation between OpenAI’s GPT-Live-1 voice model and Dessa made Dessa come across as more natural, according to the website.
On the website, the Dessa team describes the model as a “true speech-to-speech model, and everything it knows came from human speech,” and says it can run locally on a MacBook Pro M4 Max “or better.” The team said it built Dessa for “curiosity (aka fun), better understanding between humans and AI, and safety.”
In explaining the safety rationale, the team said current voice AI “flattens what we say into text and can’t keep up with human timing” and remains “nearly deaf to the emotion in our voices.” It added that a voice AI could keep going while a person is in serious pain, calling that “the safety problem” and describing Dessa as “a step toward fixing it.” The team also compared the training approach to how babies learn to talk without seeing words written down, saying a model can learn that way too but “it’s just very hard.”
The Dessa team said it will give “a small group of AI safety researchers” early access to the model.
Editor's Summary
Block has announced Dessa, an AI model trained only on human conversation and presented as a speech-to-speech system that can run locally. Sample conversations show natural exchanges alongside awkward moments, while Block frames the work around understanding, safety and curiosity. A small group of AI safety researchers will get early access.