Dessa: Speech-Only AI Learns to Converse Without Text

Block has released Dessa, a full-duplex speech-to-speech AI model trained exclusively on human conversation audio—no text, no pretrained models. The model listens and speaks simultaneously, knows when to interject or backchannel, and performs inference faster than real time on a MacBook Pro M4 Max or better. Unlike conventional voice AI that converts speech to text and discards information in the process, Dessa preserves emotion, sarcasm, laughter, and natural speech patterns.

Training exclusively from audio forces the model to learn accents, breathing, emotional cadence, and conversational timing—elements that text-based systems cannot capture. Dessa was trained on a single NVIDIA GB200 NVL72. The research team at Block, led by former members of the Toronto-based startup Dessa, contends this approach produces fundamentally different behavior from text-trained language models.

Block frames the work as addressing interface design and AI safety. The team argues current voice systems cannot detect emotion in speech, creating potential risks when users are distressed. Early access is being offered to AI safety researchers to identify unforeseen risks.