Block has announced the Block Dessa voice model, a speech-to-speech system the company says it trained from scratch using recordings of human conversations rather than text or pretrained models. The October 23 announcement comes with a familiar catch: the public can hear a few demonstrations, but cannot yet download the model, call an API, or check independent benchmark results.
Dessa lead Ragavan Thurairatnam described the system as a full-duplex speech-to-speech model in a post highlighted by The Verge. Block says the model can run on consumer GPUs. A separate report citing Block's launch material says the company claims faster-than-real-time local inference on a MacBook Pro M4 Max or newer. Those are Block's claims, not independently reported measurements.
What does full-duplex mean for Block's Dessa voice model?
In this context, full-duplex means the system is designed to listen while it speaks, rather than rigidly waiting for one speaker to finish before responding. Block says Dessa can interrupt and offer brief acknowledgments during a conversation, known as backchannels.
That approach is already an active area of voice-AI research. A May paper on spoken-language models describes human dialogue as full of pauses, overlap, interruptions and short responses such as “mm-hmm.” It contrasts concurrent speech systems with the common chained setup of voice-activity detection, automatic speech recognition, a language model and text-to-speech generation. ByteDance has also announced its own full-duplex voice model, Seeduplex, which it says is built to manage pauses and interruptions.
Block's unusual training claim is that Dessa learned only from conversational audio, with no text in its training process and no pretrained model underneath it. The available material does not establish how the conversation recordings were obtained, what languages they cover, or how the system performs across tasks. Those omissions matter more than a glossy demo reel.
What do the public demonstrations show?
The evidence so far consists of shared recordings. In one conversation highlighted by The Verge, Dessa initially calls its human interlocutor “Ronaldo,” before the exchange continues in a mostly natural rhythm. Another posted recording has the model playing a spoken guessing game. They show the system producing responsive dialogue, but they do not amount to controlled testing.
Block has not released standard benchmark results, and the supplied reporting identifies no third-party evaluation. There are also no public model weights, downloadable software or API. Early access is limited to a small group of AI-safety researchers, according to reporting that cites Block's materials.
For now, Dessa is an announcement and a set of demonstrations. Whether its speech-only training and concurrent conversation behavior hold up beyond those recordings will require public access or outside testing.
This story draws on original reporting from The Verge.
