Block presents Dessa, a speech AI model trained on human audio
Block, the company behind Square and Cash App, has introduced Dessa, a speech model that processes audio directly to recognize emotions, sarcasm and pauses in real time. The model is not yet publicly available.
Block, the company behind Square and Cash App, has introduced Dessa, a new speech model that works directly with audio rather than first converting spoken language into text. The model is designed to recognize emotions, sarcasm and pauses in real time, elements that text-based systems tend to lose when speech is transcribed.
The central difference lies in how the model handles its input. Most conventional voice systems transcribe speech into written words before analyzing meaning, and the transcript discards much of what a listener would pick up on. Tone, hesitation, emphasis and timing often carry as much meaning as the words themselves. Dessa skips the intermediate transcription step and processes the sound itself, which allows it to register these vocal cues. According to the description of the project, the model was trained entirely on human speech, which is the basis for its focus on how something is said as well as what is said.
Dessa runs locally on a MacBook Pro M4 Max, an indication that the system is built to operate on a high-end consumer laptop rather than relying solely on large cloud data centers. Local operation can mean faster responses and that audio does not have to leave the user's device, although the source material does not describe how Block handles data storage, retention or privacy in more detail.
Dessa is not yet publicly available. The source material gives no release date, no pricing, and no indication of whether the model will be offered to outside developers or integrated into Block's own products such as Square or Cash App. Nor does it include independent evaluations, benchmark results or comments from outside researchers, so the claims about recognizing emotion and sarcasm remain the company's own description for now.
What Block has disclosed so far is the technical premise: a speech model that listens rather than reads. The approach reflects a broader question in voice technology about whether machines can understand the affective layer of human communication, a layer that written transcripts flatten. Until the model is released or tested beyond the company's announcement, how well it performs on that task will remain open.
25
