Timeline
- 7.0
Block’s new voice model sounds eerily human
Block, Jack Dorsey's company, released a new voice model from its research lab that it says runs locally on a MacBook M4 Max. The company says the model, introduced as Dessa, produces human-like expressions such as laughs, pauses and cries, without cloud inference. Block presents local execution as the model's defining trait, pointing to audio generation on consumer hardware rather than through cloud inference. The company did not disclose benchmarks, licensing terms, parameter counts or a release date alongside the announcement. According to Block, Dessa is a full-duplex speech model trained only on speech, rather than on transcribed text paired with audio. The disclosed material centers on demo samples; third-party coverage notes that benchmarks and access status remain undisclosed.
- 7.0
Dessa: the most authentic full-duplex speech model ever built
Block announced Dessa, a full-duplex speech model that can listen and speak at the same time, describing it as the most authentic such model built to date. The company says it runs locally on a MacBook Pro M4 Max or better and was trained on a single NVIDIA GB200 NVL72 system. If the claims hold, combining natural interruption handling and low latency with local inference on a consumer laptop would mark a meaningful deployment step for voice AI, moving real-time spoken agents off cloud infrastructure. The release is a vendor announcement and has not been backed by third-party benchmarks. Block presents the model as both highly expressive and low-latency, and highlights the training efficiency of using one GB200 NVL72 rack. No independent latency, quality or hardware-requirement measurements were provided with the announcement.