Block Unveils a Locally Runable Full-Duplex Speech Model, but Did Not Release Its Weights

Block's AI lab has released Dessa, a full-duplex speech-to-speech model. According to coverage by Michal Sutter from MarkTechPost, it is trained exclusively on human conversational audio, and can run locally on a MacBook Pro M4 Max. Block has not announced the model size, evaluation results, nor released its weights.

To add necessary context: Full-duplex speech models are built to handle concurrent listening and speaking in natural conversations, including scenarios where users interject, pause, give backchannel feedback, or where two people's voices overlap. The system does not need to wait for the other party to finish speaking completely before it responds. Traditional voice assistants work more like walkie-talkies, returning a response only after they finish processing an entire segment of speech. Dessa's key highlight is that its training signals are sourced directly from raw human conversational audio, without going through text transcription pipelines. This approach can better match the rhythm and noise profile of real-world conversations, and may also avoid performance degradation caused by text annotation processes.

The confirmation that it runs locally on MacBook Pro M4 Max is not a trivial detail. Local operation leaves plenty of room for improvements on privacy protection and inference latency, and it also proves that the model scale or inference optimization is at least tailored for single consumer device environments. But being able to run does not equal being practically usable. Without public parameter count, we cannot tell how small the model actually is; without public benchmarks, we cannot compare it against similar models in the category; without public weights, developers cannot reproduce its performance. The only facts we can confirm right now are the official release itself and its supported operation scenario.

The information provided in this release is as thin as a draft model card. For researchers, it puts forward a new viable direction: training full-duplex speech models using conversational audio only. For ordinary users, it demonstrates the upcoming feasibility of local speech interaction. For developers, the most critical supporting evidence is still not available. The name of the model does not matter: whether it can respond properly to interruptions, process overlapping speech accurately, and what its actual latency is, are the factors that decide if it is a real, functional breakthrough.

Original article link: Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Speech, by Michal Sutter.