Block releases research preview of the Dessa voice model
BIBIBI
AT A GLANCE
Block has released the Dessa voice model, saying it was trained solely on recordings of human conversations and can run locally on certain MacBook Pro models.
Article
Block launches Dessa: It can learn to converse simply by listening to people talk and can run locally on a MacBook
Beating News AI flash, Block, owned by Jack Dorsey, has released the voice model Dessa.
Its most distinctive feature is that it was trained from scratch entirely on recordings of human conversations, without using text data or any pretrained language model.
Dessa directly receives and generates speech and supports full-duplex conversation—listening and speaking at the same time. It can naturally respond before the other person has finished speaking, and laugh out loud when it hears something funny. In the official demonstration, when someone suddenly forgot a movie title, it proactively filled it in; upon hearing about an anniversary surprise, it exclaimed and responded in a tearful voice.
Block said that Dessa uses just system for NVIDIA GB200 NVL72 training, equivalent to an entire rack of 72 GPUs Blackwell GPU. The model's inference speed is faster than real-time playback and it can run locally on a MacBook Pro equipped with an M4 Max or higher-spec chip.
Block has not yet disclosed the model's parameter count, standard benchmark results, or weights, and has provided early access only to a small number of AI security researchers.
Original link https://m.theblockbeats.info/flash/373568
Key points
01
Block has released the Dessa voice model.
02
Dessa was reportedly trained from scratch entirely on recordings of human conversations, without using text data or pretrained language models.
03
Dessa directly receives and generates speech and supports full-duplex conversation.
04
Block says Dessa uses system for NVIDIA GB200 NVL72 training.
05
Dessa can run locally on a MacBook Pro equipped with an M4 Max or higher-spec chip.
06
Block has not disclosed Dessa's model parameter count, standard benchmark results, or weights, and has provided early access only to a small number of AI security researchers.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Dessa is a model that directly processes speech without first converting it to text. It can listen and speak at the same time and attempts to respond naturally based on the conversation.
Why it matters to readers
If the related claims hold true, Dessa demonstrates a direction for training voice models using only conversation recordings and running such models locally on personal computers.
Running locally means speech processing may be performed directly on one's own computer, but specific speed, performance, and scope of applicability still require confirmation through test data.
Risks and unknowns
The model parameter count, standard benchmark results, and weights have not yet been disclosed.
Currently, only a small number of AI security researchers have received early access, and publicly verifiable information is limited.
The specific hardware requirements and performance data for running locally on a MacBook have not yet been disclosed.
Some information about the training method, interaction capabilities, and runtime performance comes from Block's statements and remains to be independently reproduced.
Related Developments
Loading event timeline…
Related concepts
Full-duplex conversation
This term is not in the glossary yet. Browse related concepts in the glossary.
