Block introduces Dessa, a speech-to-speech model trained without text

Block introduces Dessa, a speech-to-speech model trained without text

Block has unveiled Dessa, a full-duplex AI voice model that the company says was trained from scratch entirely on human conversations. It uses no text or pretrained models, according to the announcement. Dessa processes audio directly and can listen and speak simultaneously. Block says it recognizes and expresses emotional cues, including laughter, sarcasm, frustration and sadness. The company reports training it on one NVIDIA GB200 NVL72 system and is giving a small group of AI safety researchers early access.

The market significance lies in how conversational models are built. Dessa belongs in the foundation-model landscape because Block trained the underlying speech model, rather than adding a voice interface to an existing model. Direct audio processing could preserve tone and timing that speech-to-text conversion can lose. Block also says Dessa learned language from hundreds of times fewer words than LLMs, while acknowledging their broader tasks. That comparison raises a data-efficiency question, but it does not establish equivalent capabilities or lower total training costs. The announcement provides no independent evaluation.

For builders, the concrete diligence task is to test conversational control: interruptions, response timing, emotional expression and requested volume changes. Block claims Dessa can produce some behaviors without explicit training for them. Those claims need repeatable tests before they support a deployment decision. Early safety access creates a route for outside scrutiny, but it supplies no findings yet. Investors should assess the training approach separately from product readiness; the article gives no pricing, customer adoption or commercial availability details.

#Block #Dessa #VoiceAI #FoundationModels #SpeechToSpeech