Block Unveils Dessa, a Speech-Only Local Voice Model | Zeniteq
Loading
Loading article
Block Unveils Dessa, a Speech-Only Local Voice Model
Block says Dessa learns entirely from human conversation, listens while speaking, and runs on a high-end MacBook, but access remains limited to selected researchers.
Block has unveiled Dessa, a voice AI model trained entirely from human conversation without text or pretrained language models. The company claims Dessa listens and speaks simultaneously, runs locally on a MacBook Pro M4 Max or better, and was developed by Block's research group under cofounder Jack Dorsey.
Block is offering early access to selected AI safety researchers through its official Dessa page, with no general public release date announced. The announcement includes conversation recordings demonstrating expressive delivery and full-duplex capabilities, though independent performance evaluation and detailed hardware requirements remain unpublished.
Block has unveiled Dessa, a voice AI model it says was trained from scratch on human conversation, without text or a pretrained language model. According to Block’s announcement, Dessa can listen and speak simultaneously and run locally on a MacBook Pro M4 Max or better.
Dessa lead Ragavan Thurairatnam introduced the model on October 23, 2026. The announcement is a research reveal with selected conversation recordings and an invitation for safety researchers, not a generally available model release.
Three separate claims make Dessa interesting: how it learned, how it handles live conversation, and where it can run. Each requires different evidence. A convincing conversation recording can illustrate timing and expression; verifying the training recipe or reproducing local performance requires additional disclosures and access.
Speech-Only Training Without a Pretrained LLM
Block describes Dessa’s interface in five words: “Audio in, audio out.” The company says everything the model knows came from human speech, with no text or pretrained models used in training.
That distinguishes the claimed approach from a familiar voice-assistant pipeline: transcribe speech, send the text to an LLM, then synthesize its written response. Dessa is presented as learning conversation directly from sound, including both the content of speech and its delivery.
The distinction matters because words alone leave out information. A pause can signal hesitation. Stress can change the meaning of a sentence. Laughter during someone else’s sentence can be an immediate response, not an event to append after a transcript is complete. A system that retains acoustic information has an opportunity to use those signals.
Block’s announcement describes a training approach, however, without providing a detailed internal architecture. It leaves open how Dessa represents audio internally, how its components are organized, and how training objectives produce its conversational behavior. Speech-only training should not be interpreted as proof that the model uses no intermediate representations.
Block also says Dessa learned language from hundreds of times fewer words than LLMs. That is a vendor comparison without published dataset sizes or a matched evaluation. Spoken conversation and a broad text-training corpus differ in content, repetition and task coverage, so the claim cannot yet support a like-for-like efficiency ranking.
My view is that the training claim is the announcement’s most consequential part. If independently substantiated, it would offer useful evidence about what AI models can learn from conversation alone. The expressive recordings are interesting, but they cannot verify what went into training.
Work with Zeniteq
Promote your app with us
Reach readers exploring AI tools and workflows. Discover sponsorship and advertising options for your app.
Full duplex means Dessa can listen and speak at the same time, according to Block. A turn-based assistant typically waits for a person to finish before responding. A full-duplex system can continue receiving speech during its own output.
That opens up several conversational behaviors:
Backchanneling: Brief acknowledgments that show the system is following along.
Overlapping responses: Reacting before the other speaker has finished.
Interruption handling: Detecting speech while talking and deciding whether to stop, continue or respond.
Block describes recordings in which Dessa laughs mid-sentence, completes a person’s thought and gets louder when asked. The company attributes these behaviors to learning from conversation.
These are useful distinctions because conversation quality depends on timing as well as wording. An acknowledgment delivered a second too late can feel disconnected; an interruption at the wrong moment can prevent someone from finishing an important request.
Block also claims inference runs faster than real time. That concerns processing speed relative to the duration of audio. It is not an end-to-end response-latency measurement. Buffering, audio capture, decisions about when to speak and output playback can all affect the delay a person experiences.
I’d want measured interruption and response timing before judging how natural Dessa feels across ordinary conversations. Fluent delivery is one component; consistently yielding the floor is another.
Local Inference and Training Use Different Hardware
Block names two very different hardware targets: a MacBook Pro M4 Max or better for inference, and one NVIDIA GB200 NVL72 system for training.
Stage
What Block says
What remains unspecified
Local inference
Runs on a MacBook Pro M4 Max or better
Memory configuration, numerical precision, software setup and measured performance
Processing speed
Inference runs faster than real time
End-to-end latency and test conditions
Training
Used one NVIDIA GB200 NVL72 system
Training duration, utilization and total compute or cost
The official page supplies those hardware claims but no reproducible deployment recipe. It also omits the parameter count and downloadable weights.
A MacBook Pro M4 Max is a high-end consumer machine. Naming that chip family is useful, but buyers still need to know the required memory configuration and whether performance changes under sustained use. Block’s wording gives prospective researchers a hardware target, not enough information to select a configuration confidently.
The training claim needs similar care. One GB200 NVL72 system is a system count, not a claim that training used a single GPU. Without runtime and workload details, that count cannot establish total compute consumption, energy use or training cost.
Local execution could be valuable for privacy-sensitive applications because audio processing could remain on the device. That benefit depends on the surrounding application: where recordings are stored, what telemetry leaves the machine and whether other services are involved. Local model inference alone is insufficient to establish those privacy properties.
For developers, the practical implication is straightforward: Dessa has a stated local hardware target, but no public installation path yet. There is no supported setup procedure to reproduce from this announcement.
The Recordings Show Expression, Not Reliability
Block’s launch page contains several selected conversations. One involves a person struggling to remember a movie title; another presents an anniversary surprise with an emotionally delivered response. Other examples emphasize varied intonation, speed and life-advice dialogue.
The company says the recordings are live interactions, that nobody steers Dessa, and that subtitles were added afterward. Those descriptions are Block’s account of how the demonstrations were produced.
TechCrunch’s October 24 report also highlighted examples featuring changing voices, cadences and emotional delivery. Its coverage helps document the public reveal, but it is not an independent benchmark of the model.
No independent performance evaluation appears in the supplied reporting. Selected clips cannot reveal how frequently the system produces a successful exchange, how much surrounding conversation was needed, or how often it misunderstands someone.
The movie-title example illustrates the distinction particularly well. Retrieving a fact in one conversation is evidence of that demonstrated behavior, assuming the recording is presented accurately. It offers little basis for estimating factual accuracy across unfamiliar subjects.
Emotional delivery raises a separate question. A voice that sounds grateful or distressed demonstrates expressive output; it does not reveal whether the model correctly understood the other person’s emotional state. I’d keep those capabilities separate when assessing it. Persuasive expression can make an answer feel more trustworthy without making its contents more accurate.
How Safety Researchers Can Request Early Access
Block says it is giving a small group of AI safety researchers early access to investigate risks the team has not anticipated. Its research group is part of Block and reports directly to cofounder Jack Dorsey.
Researchers should go to the official Dessa page, find the section addressed to safety researchers and use its contact link. The page also has a general waitlist, which is separate from the targeted research invitation.
Block has not published selection criteria, a general-release date or API pricing. TechCrunch reported that no public access was available when its article appeared on October 24.
A useful research request could explain the proposed evaluation, the behaviors being investigated and how findings would be documented. That is a practical suggestion, not a published Block application requirement.
Dessa’s claimed design suggests several worthwhile evaluation areas:
Interruption and stopping behavior: Whether the model yields appropriately when someone interrupts, including during urgent speech.
Emotional interpretation: How reliably it distinguishes distress, sarcasm and ordinary variation in delivery.
Trust and disclosure: Whether highly expressive speech makes factual errors harder for listeners to recognize.
These are research priorities inferred from the design, not documented failures. Block itself frames sensitivity to vocal emotion as a safety issue, arguing that an assistant could miss important context if it cannot hear distress.
The next useful evidence would be reproducible hardware requirements, measured conversational timing and evaluations that include unsuccessful interactions. Until access broadens, safety researchers can request participation; other readers have Block’s technical claims and selected recordings to examine.
Frequently Asked Questions
3 questions
1
Is Dessa Available to Download?
Block has not published downloadable weights or general public access. It offers a waitlist and a separate early-access invitation for selected AI safety researchers.
2
What Hardware Does Dessa Need?
Block says inference runs locally on a MacBook Pro M4 Max or better. It has not specified the required memory configuration or published a reproducible setup.
3
Does Dessa Use a Pretrained LLM?
Block says Dessa was trained from scratch using only human conversation, without text or pretrained models. Detailed architecture and independent verification of that training claim have not been published in the supplied evidence.
