Block, the company behind Square and Cash App, has unveiled Dessa, a speech model trained entirely on human speech with no text data. The system processes audio directly instead of converting it to text first, which allows it to detect emotions, sarcasm, laughter and pauses in real time. For businesses deploying voice agents, this matters because current voice systems routinely miss the cues that shape customer intent.
Dessa design and current testing stage
Dessa was built from scratch as an audio-native model, according to Block. Conventional voice pipelines transcribe speech into text and then run language processing on that transcript, discarding tone, timing and vocal expression. Dessa avoids that intermediate step and works with sound itself, preserving signals that text-based models cannot see. The difference is central to tasks such as support calls and sales conversations.
The model supports full-duplex communication, meaning it can listen and speak at the same time. That architecture mirrors natural dialogue, where interruptions, backchannel responses and overlapping speech carry meaning. Block also reports modest infrastructure demands: Dessa runs locally on a MacBook Pro M4 Max and was trained on a single NVIDIA GB200 NVL72. Local execution reduces latency and keeps voice data on the device.
Block cites three motives for the project: curiosity, interest in better human-AI communication, and safety concerns. The company argues that voice AI which ignores emotional cues can misread users and respond inappropriately. Dessa is not publicly available yet, and the team is now seeking AI security researchers to probe the model for unexpected risks. Interested specialists can apply through a signup form.
What voice-native models mean for business
For companies using voice in sales and service, audio-native processing could change qualification and routing. An agent that hears hesitation, irritation or sarcasm can adjust its script, escalate earlier or offer a human handoff before conflict grows. Small firms could add this sensitivity without large server fleets if local deployment holds, while large contact centers could cut transcription delays and handle interruptions more smoothly.
At the same time, buyers should separate expressive delivery from answer quality. Earlier efforts show the gap: Hume AI launched EVI in March 2024, Kyutai released Moshi in September 2024, LAION has pursued open-source empathetic synthesis since late 2023, and OpenAI updated ChatGPT Advanced Voice Mode in June 2025 with better intonation and pausing. A natural-sounding voice does not guarantee sound reasoning, and the ability to reason while responding affects results more directly.
The marker to watch is the outcome of Dessa security testing and any limited release that follows. If external researchers validate safe handling of emotional signals and Block moves toward partner access, voice-native models will merit pilots in support and collections. Until then, emotion detection remains a promising input, not a proven conversion driver.
