Block unveils Dessa, a full-duplex voice model trained only on speech

Block’s AI research lab has introduced Dessa, a full-duplex voice model trained exclusively on human conversational speech data. The company said the model does not rely on text data or any pretrained model, setting it apart from systems that use text as an intermediate layer. Dessa uses a speech-to-speech architecture, allowing it to listen and speak at the same time. According to the announcement cited by Techub News, the model supports more natural interactions such as real-time interruption and laughing in the middle of a sentence. For training, Dessa was built on a single NVIDIA GB200 NVL72 rack equipped with 72 GPUs. On the inference side, Block said the model can run locally on a MacBook Pro M4 Max or higher-spec devices, with performance faster than real time. The model is still in a research preview stage. Block has not released model weights, a paper, an API, or pricing information.