Block Releases Local-Run Full-Duplex Model Dessa Trained on Pure Voice
Block has released the voice model Dessa, claiming it was trained from scratch using pure human dialogue audio, without text or pre-trained models involved.
The model features a full-duplex voice-to-voice architecture, capable of listening and speaking simultaneously, inserting dialogue, emitting backchannel sounds, and natively processing emotions, sarcasm, laughter, and sighs.
Training was completed on a single NVIDIA GB200 NVL72 rack, with inference speed exceeding real-time, and it can run locally on MacBook Pro M4 Max or higher configuration devices.
Currently, early access is provided only to a select few AI safety researchers, with a public waiting list open, but no public API, weights, or pricing available yet.
The Block team consists of former members of the Toronto startup Dessa, reporting directly to Jack Dorsey, and claims this is the first known AI capable of real-time conversation learned solely from human speech.
The company emphasizes that voice is the most natural way to interact with intelligence, as most current voice AIs lose emotional and rhythmic information due to text conversion.
Source: Public Information
ABAB AI Insight
Block previously acquired the Toronto deep learning startup Dessa, whose team has been building models and conducting generative speech research for over eight years since 2012. This time, they launched Dessa in their internal lab using a pure audio path, continuing their previous work on voice cloning and synthesis.
In terms of capital and technology, the company concentrated training power on a single GB200 rack, avoiding large-scale text pre-training clusters, motivated by the exploration of infant-like pure auditory learning and achieving local full-duplex inference, reducing reliance on cloud text pipelines.
This contrasts with early pure voice synthesis and current multimodal voice model paths, as it is currently transitioning from research previews of voice-native models to safety assessments, where coding and dialogue applications have a text-dominated foundation, while the pure audio path remains in its early stages.
Essentially, it is a technological replacement: bypassing the text intermediary layer to learn emotions and timing directly from speech, forcing voice interaction to shift from a transcription-inference-synthesis chain to a native full-duplex architecture, otherwise emotions and natural rhythms will continue to be compressed.
ABAB News · Cognitive Laws
Pure auditory learning can bypass the emotional layer lost in text
Full-duplex transforms dialogue from turn-taking to simultaneous
Local small models prove that voice-native does not need to rely on giant clusters