Block unveiled Dessa on October 23, 2026, introducing a research-preview voice model that it says learned entirely from human conversation, without text or pretrained models. Dessa can listen while speaking and run locally on a MacBook Pro M4 Max or better, according to Block’s announcement. TechCrunch’s October 24 report places the announcement the previous day.
The published recordings emphasize laughter, interruptions, pauses and emotional delivery. Block says Dessa decides what to say, when to speak and how to sound, with subtitles added afterward.
The access story is narrower than the technical pitch. Block offers demo recordings, a waitlist and early access for a small group of AI-safety researchers. It has not announced a public API, downloadable weights, pricing or a general-availability date. Local execution is a claimed capability, not something readers can currently download and try.
Dessa’s announcement combines three important properties: speech-only training, simultaneous listening and speaking, and local inference. Separating them makes it easier to understand both the research contribution and what remains unproven.
Speech-Only Training and Full-Duplex Conversation Are Different
Block describes Dessa as an audio-in, audio-out model trained from scratch on human conversation. Its language and factual knowledge come from speech, the company says, without a pretrained language model supplying a foundation.
That training approach is distinct from its conversational interface:
| Property | What Block says Dessa does | Why it matters |
|---|---|---|
| Speech-only training | Learns from human conversation without text or pretrained models | Makes spoken interaction the source of its language, knowledge and conversational behavior |
| Speech-to-speech operation | Accepts audio and generates audio | Lets the system work with vocal information beyond a written transcript |
| Full-duplex interaction | Listens and speaks simultaneously | Allows overlap, interruptions and brief acknowledgments during another speaker’s turn |
| Local inference | Runs on a MacBook Pro M4 Max or better | Makes on-device execution possible on the hardware Block identifies |
These are separate design choices. A voice system could support overlapping conversation while drawing on text-based training. Another could process speech directly but still require users to finish before answering. Local execution says where a model runs, not what it learned or how it handles conversation.
Block’s description of Dessa brings these properties together, but it does not publish enough architectural detail to explain precisely how the model represents audio or coordinates listening and generation.
My view is that speech-only learning is the most consequential research claim here. Natural-sounding output is an immediately noticeable feature, but learning language and conversational behavior without a text-trained starting point raises a different set of questions about what AI models can learn from sound.
The Recordings Focus on Timing and Vocal Expression
Block’s examples concentrate on details that ordinary transcripts capture poorly. Its descriptions include:
- Laughing while a person is still speaking.
- Completing a sentence when someone cannot remember a movie title.
- Increasing volume when asked to speak up.
- Producing a surprised gasp and emotionally expressive response.
- Giving short acknowledgments while listening.
These behaviors involve prosody, the rhythm, emphasis, pitch and pacing of speech. A transcript can preserve the words in a sentence while losing the difference between a sincere answer, a sarcastic reply and an exhausted acknowledgment.
Timing carries information too. A brief acknowledgment during someone’s explanation can signal attention without taking over the conversation. A pause can invite the other person to continue. Full-duplex operation gives a model the opportunity to respond to those cues while audio is still arriving.
Block says the recordings are live conversations without someone steering Dessa’s responses. Those are company-provided demonstrations, however, and this article assesses the published material, not hands-on testing. The recordings illustrate behaviors Block wants to highlight; they provide no independent measurement of how reliably those behaviors occur across users or situations. See the official demo page.
Expressive delivery also needs careful interpretation. A voice that sounds sympathetic can make an interaction easier to follow, but it does not by itself establish accurate recognition of a person’s emotional state. I’d want evaluation of whether Dessa responds appropriately to distress, sarcasm and ambiguity, including cases where its first interpretation is wrong.
Block calls Dessa the “most authentic full-duplex speech model ever built” and presents its speech-only conversational training as a first, qualified by the company’s knowledge. The announcement supplies no public benchmark supporting either a competitive ranking or that priority claim.
Local Inference Has a Specific Hardware Requirement
Block names different hardware for training and running Dessa:
| Stage | Hardware disclosed | What the claim covers |
|---|---|---|
| Training | One NVIDIA GB200 NVL72 system | Infrastructure used to build the model |
| Inference | MacBook Pro M4 Max or better | Hardware on which Block says the trained model runs locally |
The training claim should not be shortened to one GPU. Block names a GB200 NVL72 system, and TechCrunch describes the infrastructure as a single NVIDIA GPU rack. A single system or rack is a very different resource claim from a single accelerator.
Training and inference also have different demands. Training builds the model by repeatedly processing examples and updating its parameters. Inference uses the resulting model to generate responses. Dessa’s reported Mac runtime therefore says little about the resources needed to reproduce its training.
For local use, the named Mac is a useful starting point, but not a complete specification. Block does not provide the required memory configuration, model size, power consumption or a list of equivalent supported machines. “Or better” leaves potential users without a precise compatibility boundary.
Faster Than Real Time Is Not a Latency Benchmark
Block also says Dessa’s inference runs faster than real time. In practical terms, that describes processing throughput sufficient to keep pace with a live audio stream. It does not specify the delay between a person speaking and hearing Dessa respond.
A conversational system can process audio quickly while still waiting to gather context, buffering input or delaying its decision to speak. Audio playback adds another part of the interaction. None of those delays can be calculated from the announcement’s speed claim.
The important missing measurement is end-to-end conversational latency under a stated hardware and software configuration. Interruption handling deserves measurement too: how quickly does the model stop, acknowledge or adapt when the other speaker cuts in?
Local execution could reduce dependence on a remote server and create opportunities to keep conversational processing on-device. Those are practical implications of the design, not announced privacy guarantees. The preview does not establish how a future application would handle recording, storage or network access.
Frequently Asked Questions
3 questions
1Can I Download Dessa or Use Its API?
Block has not announced downloadable weights or a public API. The preview offers recordings, a waitlist and limited early access for AI-safety researchers.
