Block has introduced Dessa, a speech model it says can listen and answer faster than real time on an M4 Max MacBook Pro, bringing local operation to laptop hardware.
A speech-to-speech system takes spoken sound as input and produces an audible response. Block says Dessa can start responding while incoming audio is still being processed, allowing listening and speaking to overlap.
Block’s AI research lab developed the model from recorded conversations. The company says training used neither written text nor an existing pretrained model, and relied on one NVIDIA GB200 NVL72 system.
Block names a MacBook Pro with M4 Max, or a computer with higher specifications, for running the model locally. The NVIDIA system was used for training, while the MacBook is the company’s example of a device for local use.
Block has released audio demonstrations and opened access to a small number of AI safety researchers. Dessa remains a research preview and is not generally available.
