Voice assistants always fall short: they only speak after you finish, cut off instantly when you interrupt, and sound completely flat. Block (founded by Jack Dorsey, parent company of Square) unveiled its answer on October 24 Dessa— A conversational model trained from scratch exclusively on human speech. No text, no transcription, no text-based pretraining—not a single word was ever written down.
This approach is fundamentally different from the path we’re familiar with.ChatGPTGemini’s voice capabilities are essentially a three-stage pipeline: “speech recognition → text-based large language model → speech synthesis,” whereas Dessa has only one stage:Audio in, audio outBlock’s explanation is straightforward—babies learn to speak without ever seeing written words, and they wanted to see whether a model could learn the same way.
What exactly is Dessa?
The official definition is “a true speech-to-speech model whose entire knowledge comes from human speech.” Key features include:
- Full-duplex— It can listen while speaking, inserting interjections like “hmm” or laughter mid-utterance, or even interrupting to correct itself—just like humans—rather than waiting for you to finish before processing.
- Latency under 200 milliseconds— This figure approaches natural human conversational pauses; delays exceeding 500 milliseconds make users distinctly feel “the machine is waiting.”
- Conversational memory— Within a single session, it remembers prior exchanges and maintains contextual continuity, rather than resetting to zero with each turn.
- Run locally— Officially stated to run locally on a MacBook Pro M4 Max or more powerful hardware—no cloud connection required.
- Native understanding of paralinguistic information— Tone, emotion, sarcasm, laughter, pauses—elements that vanish entirely upon transcription into text—are directly perceived by the model.
Training data volume is just one-thousandth that of text-based models.
This is, to me, the most intriguing part. Text-based large models are typically trained on several trillion tokens, whereas Dessa’s training volume, per official disclosure, isIn the range of thousands of hours of speechBlock’s team claims that one hour of speech is roughly equivalent to a book’s worth of text, and it would take ten billion hours to reach the trillion-word scale—yet they’ve slashed several zeros off that figure, betting on the premise that “speech itself carries higher information density.”
In other words, Dessa is areverse experiment: while everyone else is bolting speech capabilities onto text models, Block is attempting to prove that “learning to speak only” is viable. If successful, it implies low-resource languages, dialects, and colloquial expressions could be covered without massive textual corpora—a prospect with considerable implications for Chinese dialect scenarios.
Demo performance: natural, but clearly immature
In several audio clips released officially, Dessa comes across as “an earnest but somewhat unpolished new friend”: while discussing Nolan films, it abruptly shifts to Bryan Singer’s Superman; when told it will be taken to Paris, it replies with something resembling “a tearful thank-you”; and it mistakenly identifies the speaker as someone named “Ronaldo” at the start of a game.
Another clip features Dessa conversing with OpenAI’s GPT-Live-1 voice model; The Verge judges Dessa as sounding more natural. Yet, in my view, such officially curated samples are best taken with a grain of salt—real-world experience requires hands-on use, and right now youcannot access it.
Three critical limitations that must be clarified
First, it is not a product—at present.Block explicitly states that Dessa is a research project; its weights are not open-sourced and access is granted only to a small cohort of AI safety researchers for vulnerability probing. Those hoping to download and run it locally should abandon the idea.
Second, it supports English only.The speech-only training approach has no known performance metrics for Chinese—tones, erhua, and dialectal variations pose entirely new challenges for a model trained “on sound alone.” No Chinese evaluation data exists yet.
Third, hardware requirements are substantial.A MacBook Pro with M4 Max chip is the minimum, implying unified memory of at least 48 GB. This mirrors the earlier situation with local large-model inference: technically feasible, yet out of reach for most users. For insights into such barriers, see Muse Spark Open benchmark and Recursive Flint benchmark—both exemplify “plug-and-play laptops” whose entry thresholds remain prohibitive.
Where does Dessa sit within the voice AI landscape?
| Product | Technical approach | Can it be used directly? | This Site’s Rating |
|---|---|---|---|
| Block Dessa | Pure speech training, full-duplex, runs locally | Researchers only, not open to the public | Not rated yet |
| Gemini Live Avatar | Text LLM + real-time speech + video avatar | Available, supports 97 languages | 8.6 |
| ChatGPT Voice Agent | Speech-to-text + Agent execution + text-to-speech | Available, cross-app operations | 8.4 |
| GPT-6 Sol / Luna | Text flagship model; speech uses Live API | Available | 8.9 |
The conclusion is clear:Dessa may lead on the single metric of “naturalness,” but scores zero on usabilityGemini Live already enables real-time video conversations in 97 languages,ChatGPT and voice agents can check your email across apps—these are things you can use today. Dessa’s current value lies in proving the feasibility of a technical approach, not delivering a usable tool. On why voice AI has been slow to achieve the “always-on” moment, see this articleChatGPT Industry interview analysis Why this warrants close attention.
Two reasons. First,
architectural differences drive experiential differences: Today’s all voice assistants suffer “lag” due to sequential ASR→LLM→TTS latency; Dessa’s end-to-end single-model approach collapses the pipeline into one stage, theoretically approaching human responsiveness. Second,The “lag” in all current voice assistants stems from the sequential delay across ASR→LLM→TTS; Dessa’s single-model end-to-end approach compresses this pipeline into one stage, theoretically bringing it closer to human responsiveness. Second,Safety motivationBlock states clearly that teaching models to speak from acoustics is a completely new domain—so external researchers must first identify risks, including deception risks arising from outputs that “sound convincingly human but are actually machine-generated.”
As for why Jack Dorsey built this, the team cites three reasons: “curiosity (i.e., fun), improving human–AI communication, and safety.” It sounds less like a commercial plan and more like Block’s consistent ethos.
Frequently Asked Questions (FAQ)
Can Block Dessa be downloaded and used?
Not currently. Block positions it as a research project; model weights are not publicly released and access is restricted to a small group of AI safety researchers. The official website hosts only demo audio recordings, with no application portal.
Does Dessa support Chinese?
No Chinese-language capability data has been released officially; all demos are in English. Performance of pure speech-trained models on Mandarin tones and regional dialects remains unexplored, and expectations should be low in the near term.
What is a full-duplex speech model?
Full-duplex means the model can simultaneously receive and generate audio: it listens while you speak and can interject with utterances like “uh-huh” or “right” at any moment—you may even interrupt it mid-response. In contrast, half-duplex models operate strictly turn-by-turn: “you finish speaking, then it responds,” allowing only one direction at a time.
What hardware configuration is required to run Dessa?
Officially, a MacBook Pro with M4 Max or better is recommended—specifically devices with unified memory of 48 GB or higher. Even if weights are released later, most standard laptops will struggle to run it.
How does Dessa differ from Gemini Live andChatGPT speech capabilities?
The key distinction lies in training methodology: Dessa uses speech-only input without text, whereas Gemini and ChatGPT are large language models augmented with speech interfaces. In practice, Dessa feels more natural but supports only English and remains unavailable; the latter two offer broad capabilities, multilingual support, and immediate usability today.
Summarize
Dessa is a valuable proof-of-concept—not a product. It demonstrates that the path “learning dialogue solely from speech” is at least viable, while also establishing benchmarks for full-duplex operation, sub-200ms latency, and local execution. For users in China, the only actionable step now is observation—awaiting weight release or adoption of this approach by a domestic vendor.
To continuously track real-world usability and benchmark scores of new models, refer to our AI Model Library and Tool Comparison Engineor continue reading:Liquid AI LFM 2.5 Review · Amazon Bee personal AI device · All Reviews.