By ByteBot
Block, Jack Dorsey’s fintech company, announced Dessa this week — a full-duplex speech model trained from scratch on human conversation alone. No text. No pretrained models. Audio in, audio out. The demos sound convincingly human, fraud researchers are alarmed, and developers building voice agents have something genuinely new to think about.
Why Dropping Text From the Pipeline Matters
Every major voice AI product today uses text as a relay. A user speaks, the system transcribes to text, an LLM generates a text response, then a separate TTS model converts back to audio. The pipeline works, but it loses something at every handoff: tone, sarcasm, the micro-pause before an awkward sentence, the way someone’s voice tightens when they’re uncertain. None of that survives transcription.
Dessa skips the relay entirely. Block trained the model directly on recorded human conversation — no text labels, no pretrained text backbone — so the model learns what speech actually sounds like rather than what it means on a page. Emotional expression is baked into the model’s weights, not painted on afterward via voice tags. Block claims Dessa learned language understanding from hundreds of times fewer words than a comparable LLM. If independently verified, that would be a significant efficiency claim for audio-native training.
Full-Duplex and What It Means in Practice
Dessa is full-duplex: it listens and speaks simultaneously. Most voice AI uses turn-based conversation — you finish, it responds. Dessa can be interrupted mid-sentence. It backchannels naturally. In the demos Block released, it laughs in response to humor and adjusts tone to emotional cues in real time.
The infrastructure story is compact: Block trained Dessa on a single NVIDIA GB200 NVL72 rack — 72 GPUs, one rack, from scratch. The model runs locally on a MacBook Pro M4 Max. For a model making these claims, that is a small training and inference footprint. TechCrunch reported the inference runs faster than real time, so responses arrive without the pause that exposes most voice AI.
The Fraud Problem That Block Saw Coming
The demos went viral, and the backlash arrived quickly from a specific direction. Journalist Shiv Aroor: “Scam calls about to attain pandemic level.” Filmmaker Justine Bateman: “So you can trick people into lowering their guard.”
Voice is still used as authentication. Banks verify customers by voice. Corporate security lines use voice recognition. A speech model that sounds indistinguishably human — and runs locally, without an API call to an external server — creates obvious problems for any system that trusts voice as proof of identity. CoinDesk covered the scam fears in detail, noting community sentiment split 60/40 positive-to-alarmed.
Block is not dismissing this. The company is recruiting AI safety researchers as the first access group, specifically to identify deception scenarios before broader release. The right instinct — though Block announced publicly before that review finished.
What Dessa Cannot Do Yet
This is a research preview. Dessa has no tool-calling capability, cannot access external documents, and cannot query a database. For a production voice agent that books meetings, retrieves account information, or handles transactions, voice quality alone is insufficient. The model needs a brain behind the voice that can take actions. Dessa does not have that yet.
The closest comparable system is Kyutai’s Moshi — an open-weights full-duplex model with published specs (7B parameters, 200ms latency). Moshi also lacks production tooling, but at least has documented architecture. Block has published no benchmarks, no parameter count, no latency figures — only curated audio recordings Block itself selected. The community has noticed, and independent evaluation has not happened.
What Developers Should Watch
Dessa is not available for developer use today — just a waitlist. But the direction matters. Three things worth tracking:
- Independent benchmarks vs. GPT-Realtime-2 and Moshi — naturalness and latency tests will determine whether audio-native training actually outperforms cascade pipelines at scale
- Open vs. closed weights — Block has not committed to open-sourcing Dessa; with Kyutai having published Moshi, there will be community pressure
- Tool integration API — voice quality does not make a voice agent; the architecture for connecting Dessa to backend systems will define whether it becomes useful for builders
The technical bet — that audio-native training preserves conversational quality that text-based pipelines discard — is plausible, and the demos support it. Whether that translates from research preview to production-grade voice agent is the question Block has not answered yet.
