Block Unveils Dessa Full-Duplex Speech-to-Speech AI Model Trained Solely on Human Audio
What happened
Block introduced Dessa, a full-duplex speech-to-speech AI model trained from scratch exclusively on human conversation recordings without any text data or pretrained models.
The Context
- Bypassing the text layer: Standard conversational agents convert speech to text, run a language model, and convert back to audio; Dessa processes sound directly to retain pauses, emotional inflection, and interruptions.
- Local edge deployment: The model executes inference faster than real time on high-end Mac hardware, bypassing the need for remote cloud routing during voice interactions.
- Gated initial release: Block restricted early access to select AI safety researchers to evaluate synthetic speech misuse risks before releasing public benchmarks or weights.
The Number
— NVIDIA GB200 NVL72 rack was used to train the entire model from scratch, signaling that voice-native systems can emerge from compact compute footprints rather than hyper-scale clusters.
Takeaway
Watch for the release of independent benchmarks and open weights as Block moves Dessa from safety evaluation toward client-facing deployment.
News and guidance for IT pros on AI adoption.
"Enterprise-focused tips, explainers, and news for professionals."
— A47 Editor
Block Previews Dessa, an AI Voice Model Trained Without Text
Block has previewed Dessa, an AI voice model that processes audio directly and was trained entirely on human speech rather than text, enabling it to recognize emotions, sarcasm and pauses in real time. The model reportedly runs locally on a MacBook P...
Curated tech headlines including AI stories.
"Influential aggregator surfacing the day’s top tech/AI links."
— A47 Editor
Block unveils Dessa, a "true speech-to-speech" AI model trained on "nothing but human conversation" that it says runs locally on a MacBook Pro M4 Max or newer (Jay Peters/The Verge)
Block has introduced Dessa, a new AI model designed for speech-to-speech interactions, which has been trained exclusively on human conversations. This model is capable of running locally on a MacBook Pro M4 Max or newer, showcasing advancements in AI...
Daily AI news: models, tools, and policy.
"Independent outlet tracking the fast pace of AI."
— A47 Editor
Block's new AI model Dessa was trained entirely on human speech to catch emotions text-based models miss
Block has unveiled Dessa, a new AI model that processes audio directly to recognize emotions, sarcasm, and pauses in real time, trained entirely on human speech. This model operates locally on a MacBook Pro M4 Max but is not yet publicly available.