Editorial illustration for Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Speech
Block's Dessa: Full-Duplex Speech Model Trained Without Text
Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Speech
Block's AI research lab put out a speech model on October 23, 2026, that skips a step almost every voice assistant relies on. The model, called Dessa, is full-duplex, meaning it listens and talks at once, and it never touched a line of text during training. No transcripts, no pretrained language model underneath, just raw human conversation fed straight through a speech-to-speech pipeline. That's a break from the usual setup, where audio gets converted to text, run through a language model, then synthesized back into speech, a chain that tends to flatten out laughs, sighs, and the timing of a well-placed interruption.
Block trained Dessa on a single NVIDIA GB200 NVL72 rack, 72 GPUs total. For comparison, Kyutai's Moshi, another full-duplex model, used 1,016 H100 GPUs to get there. Dessa runs locally on a MacBook Pro M4 Max or better, faster than real time, according to Block.
There's no published benchmark, no paper, no weights, and no price tag yet. What Block has shown instead is a handful of demo clips, including one where Dessa holds a conversation with OpenAI's GPT-Live-1.
Most voice systems chain 3 parts: speech recognition, a text model, then speech synthesis. Block team argues that chain throws away emotion, sarcasm, laughter and sighs. Dessa hears and produces those signals directly.
Why this matters
Dessa is a research preview, not a product, and Block hasn't published benchmarks beyond "runs faster than real time on an M4 Max." That matters for anyone building voice agents right now: the interesting part isn't whether Dessa beats existing stacks on some leaderboard, it's the training approach. Skipping text and pretrained models entirely, and learning turn-taking and backchanneling straight from raw conversation, is a direct challenge to the text-first pipeline nearly every voice assistant uses today, from Siri to the various GPT-wrapped call-center bots. If audio-native training actually produces more natural interruptions and "mm-hmm" timing without a transcription layer, that's worth watching closely as latency and naturalness keep being the two things that make voice agents feel robotic.
We'd treat the local-only, Apple-silicon deployment angle as a signal too: Block seems to be testing whether speech models can live on-device rather than in a data center. Until there's a public eval or an API, this is a demo worth bookmarking, not a tool worth building on. The real test comes when Block, or someone else, tries to scale this approach past a tech demo.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Google DeepMind releases Gemini 4 Carbon; independent evaluations back gains
From Archives DeepSeek-V4 Flash RL Runs on AMD Instinct MI308X GPUs
Common Questions Answered
What makes Dessa different from traditional voice assistants in its training approach?
Dessa is trained exclusively on raw speech data without any text transcripts or pretrained language models, using a direct speech-to-speech pipeline. This contrasts with conventional voice systems that convert audio to text, process it through a language model, then synthesize speech back, which Block argues loses important vocal nuances like emotion, sarcasm, and laughter.
How does Dessa's full-duplex capability improve upon traditional voice assistant architecture?
Dessa can listen and talk simultaneously, whereas most voice systems chain together three separate components: speech recognition, a text-based language model, and speech synthesis. By processing speech directly without converting to text, Dessa preserves emotional cues, sarcasm, laughter, and sighs that are typically lost in the traditional pipeline.
What specific vocal signals does Dessa preserve that traditional voice systems lose?
According to Block's research team, the conventional three-part chain of speech recognition, text models, and synthesis throws away emotion, sarcasm, laughter, and sighs. Dessa hears and produces these signals directly because it operates on raw speech without intermediate text conversion, maintaining the full acoustic richness of human conversation.
What is the current status of Dessa and what performance metrics has Block published?
Dessa is currently a research preview rather than a commercial product, and Block has only disclosed that it runs faster than real time on an M4 Max processor. The company has not published comprehensive benchmarks, making the significance of Dessa's innovation primarily about its novel training methodology rather than performance comparisons against existing voice systems.
Why is Dessa's training approach significant for the future of voice agent development?
Dessa demonstrates that skipping text entirely and learning turn-taking and backchanneling directly from raw conversation is viable, directly challenging the text-first pipeline that nearly every current voice system relies on. This approach suggests a fundamentally different way to build voice agents that could preserve natural conversational qualities while potentially simplifying the technical architecture.
Further Reading
- Block's new voice model sounds eerily human - TechCrunch
- Jack Dorsey’s Block unveils new voice AI model that talks like a real human - The Indian Express
- Awkwardly Natural: Dessa Sounds Like Someone You Just Met - Implicator AI
- Block releases Dessa, a full-duplex speech model trained without text - Extrapolator AI
- Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Speech - MarkTechPost
