Block Unveils Dessa, a Human-Sounding Voice AI That Runs Locally

Jack Dorsey's Block releases a new voice model from its research lab called Dessa, which runs locally and sounds eerily human with expressions like laughs, pauses, and cries. The company says the model trains fully on human speech using a single NVIDIA GPU rack, and it works as a full-duplex system that can talk and listen at the same time, similar to OpenAI's GPT-Live model released earlier this year.

Typical speech models pair with a large language model for knowledge retrieval, but Block says Dessa draws its knowledge directly from conversational data. The company argues current voice AI sounds unnatural because text in the pipeline drives speech generation, and while tools like ElevenLabs v4 use voice tags to convey emotion, Block claims Dessa builds emotion directly into the model, though it remains unclear whether tags play a role.

Block's researchers share multiple demo conversations showing Dessa using different voices, cadences, and emotions, including one where the bot sounds tired when discussing work and another where it grows nearly emotional while talking. The team calls Dessa a "GPT3 moment for speech," positioning it as the first AI model capable of truly realistic conversation trained entirely on speech.

Read More at the original source →