Block unveils Dessa, a speech model trained entirely on human audio

What happened

Block unveiled Dessa, an AI model it says was trained from scratch entirely on human speech without text data, processing audio directly to capture emotions, sarcasm, laughter, and pauses in real time. It supports simultaneous listening and speaking, runs locally on a MacBook Pro M4 Max, and was trained on a single NVIDIA GB200 NVL72. Dessa is not publicly available yet; the team is seeking AI security researchers to test for unexpected risks.

Read next

• More changes in today’s briefingMistral releases improved Mistral Large 4 open weights for coding and cybersecurity

Reported by 3 sources

The story so far

Block has introduced Dessa, a speech-to-speech AI model, and is inviting a small group of safety researchers to test it for unexpected risks while public access remains unavailable. Block says the model was trained from scratch entirely on human conversation, without text data, and can run locally on a MacBook Pro M4 Max or better. Dessa can listen and speak simultaneously, with vocal expressions such as laughter, pauses and cries. Block’s shared demonstrations show mostly natural exchanges, though some moments are awkward. The early testing invitation gives safety researchers access to examine the model’s behavior before broader availability.