- Block says its Dessa model learned live conversation from human speech alone, with no text, and so can hear emotion that text-based voice AI discards.
- The evidence so far is Block's own demo clips. Dessa is not public, and Block is offering early access only to a small group of safety researchers.
- Leaders buying voice AI should ask whether a product acts on audio or on a transcript, and test answer quality rather than how natural the voice sounds.
A transcript is a summary of a conversation
Think of the minutes from a meeting. They record who said what. They do not record who hesitated, who laughed, or who sounded tired. Most voice AI works like those minutes. Block, the company behind Square and Cash App, argues that voice AI first converts speech into text. In its view, that conversion discards cues such as mood and tone.
Block's small AI research lab has announced Dessa, a model it says takes a different route. Block says it trained Dessa from scratch on human conversation only. The training material was sound, with no text and no pretrained models. "As far as we know," Block writes, Dessa is the first AI for realistic live conversation trained only on human speech.
How a model learns to talk without reading
Dessa is described as a speech-to-speech model. Sound goes in and sound comes out. Block calls it full-duplex. It can talk while the other person is still talking, so conversation is not strict turn-taking. Block says it learned when to cut in, and when to drop in a short "yeah" or "got it" to show it is following.
The hard part is that a model with no text has one signal to work from. Words, stress, tone and timing all arrive mixed together in the sound. The model has to separate them itself. A text model receives the words already spelled out.
Block's own example is the sentence "I didn't say she stole my CUDA code." Stress a different word each time and the meaning shifts. Block adds that speech never repeats exactly. Say one sentence twice, identically, and the recorded data still differs. Typed text does not vary that way. Block also points to data size: by its account, audio of a single word is thousands of times bulkier than the typed version, which makes learning from sound alone a much heavier task.
Block makes several claims about efficiency. It says Dessa needed hundreds of times fewer words of language than text-trained models. It says the model generates speech faster than real time, so users face no awkward wait. It says Dessa runs on a laptop, specifically a MacBook Pro with an M4 Max chip or better. It says training used one NVIDIA GB200 NVL72 system. The material reviewed gives no benchmarks to check any of this. It also does not state how much training data Dessa actually used.
What Block shows, and what it does not
Block published clips it describes as live recordings of people talking to Dessa. In one, the model laughs mid-sentence. In another, it finishes a sentence when a person blanks on a movie title. In a third, it gasps and answers with what Block describes as teary gratitude. Block added the subtitles afterward, so they are not part of what the model heard. One demo pairs Dessa with a third-party AI model, GPT-Live-1, with both voices AI-generated. These are the clips Block chose to publish. The material reviewed shows no independent evaluation.
Dessa is not the first attempt at emotional voice AI. The Decoder lists Hume AI's EVI from March 2024 and Kyutai's Moshi from September 2024. It adds OpenAI's reworked Advanced Voice Mode from June 2025. The Decoder cautions that sounding emotional is not the same as being emotional or intelligent. For advanced speech AI, it says, reasoning while responding matters more than how natural the voice sounds.
That is a fair test to apply to Dessa. Block's own material says little about how well the model reasons.
The decision hidden inside a voice product
This is WebPulse's argument, not Block's finding. If a product keeps the transcript, that becomes the record of the call. It can also become the material other systems learn from. Block claims that emotion lives in the audio and is lost when speech becomes text. If that claim holds, a transcript-only record leaves out part of what was said. The material reviewed shows no independent test of the claim.
Block frames the issue as one of safety. Its example is a person in serious pain while a voice AI carries on without noticing. That is the case Block makes for models that hear.
There is a second side, and it is our view, not Block's. A system that hears distress also handles a more sensitive signal. Inferring emotion from a voice is a judgment about a person. Block's material does not say how Dessa's outputs would be stored or governed.
Block has asked a small group of AI safety researchers to try Dessa early and look for dangers its own team has not thought of. The Decoder reports that the team is still looking for researchers and has posted a sign-up form. Dessa is not publicly available, and the material reviewed shows no published independent testing.
Questions to put to your team
Ask every voice vendor whether its product acts on the audio or on a transcript of it. If it is a transcript, ask which tone and timing signals are lost, and whether that matters for your customers.
Ask your legal and privacy teams what happens if a product starts inferring emotion. Where would that inference be stored, and who could see it?
When piloting any voice agent, score the quality of its answers, not how pleasant it sounds. Ask the vendor who has tested it outside the company.
A transcript is a summary of a call, not the call. The question for buyers is what their systems keep once a conversation becomes text.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Block (Square and Cash App).
