Block's Dessa is a full-duplex speech-to-speech model trained only on human audio, running locally on an M4 Max
- Block trained Dessa from scratch on human conversation audio alone, no text and no pretrained models, and calls it the first AI you can hold a live conversation with that learned only from speech.
- It is full-duplex speech-to-speech: it hears and talks at the same time, backchannels with "yeah!" and "got it!", and runs inference faster than real time locally on a MacBook Pro M4 Max or better.
- Block says training fit on a single NVIDIA GB200 NVL72, and that Dessa learned language from hundreds of times fewer words than LLMs (no benchmark numbers are published).
- The demo clips are live recordings with subtitles added afterward: Dessa laughs mid-sentence, gets louder when asked, finishes a movie title the human blanks on, and in one clip talks with GPT-Live-1, a third-party model whose developer has not reviewed or endorsed it.
- There are no weights, no API, and no pricing, only a waitlist, and the company says more recordings will be posted on X.
Hacker News 의견들
The sample on the homepage sounds awful to me. The cadence, the tonal range, the overall feel, it's so far off passing as real that I figure I must be missing something.
Have you tried the alternatives? The fluidity of conversation here is far better than any other AI voice I've heard, even if a few moments sound off.
Surprised it took this long. The speech realism is very good, but the intelligence is pretty clearly 2024-level. Next step has to be joint training on text and audio.
These duplex demos always come off as artificial. Fine, it can talk while I talk, but training it to interrupt as much as possible is the wrong goal. In the first clip the human never finishes a single thought.
OpenAI ruined advanced voice mode for me with exactly that. I'm mid-sentence and it goes "uh huh" and it kills my whole train of thought.
I get irrationally angry when actual people interrupt me. When AI support agents start doing it I'm going to lose it.
You can see cracks even in the first clip. He asks it to speak louder, it's louder for a second, then it drops back to the same level and forgets what it just said.
The wild part is training on one GB200 NVL72 while LLMs need way more compute. Does that efficiency transfer to text, or is speech just a much smaller problem?
Not if you want one model that debugs your C++ and sorts objects by shape and picks a birthday gift for your cousin. That generality is what the compute buys.
Both speakers sound unnatural. "How's the winter stuff?" is not a sentence a person says. I could buy the voice, I can't buy the conversation.
I thought it sounded genuinely natural and it freaked me out in a good way. You can tell the model has no idea what it's talking about, but the emotion in the voice is real.
That's the interesting part to me, voice carries a lot of emotional signal that text just loses. Words only cover the peaks of what we feel.
Feels like a parlor trick where the model was trained to mimic phone mannerisms. Lots of style, not much substance.
Ok, sounds cool. Now what? Open weights? An API? A listing on OpenRouter? Is this a product or a demo?
Open weights or it didn't happen.
The scam angle worries me more than the girlfriend angle. My parents would not be able to tell that voice from a real person on the phone.
The one use case I can actually picture is game NPCs that hold a plausible conversation while still getting the plot points across.