Block Unveils Dessa: A Full-Duplex Speech Model Trained Purely on Human Voice

For more than a decade, conversational artificial intelligence has relied on a three-stage cascade: speech recognition (ASR) transcribes voice into text, a large language model generates a textual answer, and a text-to-speech engine synthesises the result back into sound. While modern pipelines have reduced latency to acceptable thresholds, the fundamental flaw remains: text strips away the vast majority of human communication.

Intonation, cadence, breathing pauses, sarcastic inflection, hesitant stammering, and emotional strain are completely discarded at the transcription boundary. When an AI cannot hear how someone speaks, it cannot respond appropriately to distress, humour, or irony. Block’s Dessa tackles this limitation by abandoning text tokens entirely. Built from scratch on raw human conversation, Dessa learns language, semantics, and acoustic dynamics simultaneously from raw waveform data.

In human conversation, communication is never half-duplex. We do not wait for the other speaker to transmit an end-of-turn token before processing what was said. We nod, offer vocal acknowledgements such as “yeah”, “got it”, or “right”, and frequently interrupt or chime in mid-phrase.

Because Dessa is full-duplex, it listens continuously while generating speech. It exhibits emergent behaviours that were never explicitly programmed: it chuckles in the middle of speaking when a humorous remark lands, finishes a partner’s sentence when they hesitate, and automatically increases its vocal projection when instructed to speak louder. In live testing with third-party models like GPT-Live-1, Dessa demonstrates that conversational flow can be sustained dynamically across autonomous speech agents.

Frontier language models require trillions of tokens harvested from the entire public internet. By contrast, Dessa learned language from hundreds of times fewer spoken words than typical LLMs. Despite speech waveforms containing thousands of times more raw bytes per word than text, the research team successfully trained the entire model on a single NVIDIA GB200 NVL72 rack.

Crucially for privacy and production deployments, Dessa’s inference engine is designed to execute locally. The model runs faster than real time on Apple Silicon hardware, specifically a MacBook Pro equipped with an M4 Max chip or higher. Local execution eliminates cloud transmission latencies, safeguards private voice data from upstream surveillance, and proves that audio-native foundation models do not require hyperscale server clusters for inference.

The team behind Dessa, originally part of the Toronto-based machine learning startup acquired by Square/Block in 2020 and led by Ragavan Thurairatnam, reports directly to Block co-founder and Chairman Jack Dorsey. As the team expands the research preview, they have opened applications for AI safety researchers to examine emergent failure modes unique to speech-first models.

Traditional safety guardrails are almost universally designed around text token classifiers. A model that operates purely on raw audio requires novel interpretability tools, acoustic boundary checks, and audio alignment methods. By providing verifiable recordings and inviting external scrutiny, Block is laying the groundwork for a future where voice interactions feel human, empathetic, and fundamentally safe.