Dessa: Block's Voice Model That Mimics Human Emotion

Imagine a conversation with a voice assistant that doesn't just answer your questions, but laughs at the right moment, sighs when talking about a hard day at work, and lets its voice catch with emotion when expressing gratitude. That is precisely what Block — the company led by Jack Dorsey — is attempting to deliver with its new voice model, Dessa, developed in the company's in-house research lab and capable of running locally on consumer hardware without relying on cloud services.

What sets Dessa apart technically?

The crux of Dessa's design is that it is a speech-to-speech model trained entirely on human speech rather than on text. Most voice systems in use today operate through a three-stage pipeline: converting the user's speech into text, feeding that text to a large language model to generate a response, then converting the response back into audio. This textual intermediary is precisely why a great deal of information is lost — information that text simply cannot carry, such as tone, emotion, speaking pace, and pitch.

Block states the problem plainly in its blog post, arguing that converting speech into text means discarding a rich layer of meaning, and that the world's most powerful AI systems today "can't hear how you feel" — something the company considers an incomplete experience, and potentially a dangerous one in certain contexts.

Two-way conversation with no waiting

Dessa is a full-duplex model, meaning it can speak and listen at the same time, much like OpenAI's live voice model. This capability eliminates the rigid turn-taking that makes conversations with traditional voice assistants feel stilted: you can interrupt it, and it can chime in with a brief acknowledgement while you're speaking — exactly as happens in natural human dialogue.

More striking still, the company says the model's knowledge is derived from the conversational data it was trained on, without relying on a separate large language model to retrieve information — an architectural choice that diverges from the prevailing approach in this field.

Built-in emotion instead of artificial "tags"

Many current voice generation models, such as certain releases from ElevenLabs and Fish Audio, rely on text tags to steer emotion: one tag for laughter, another for whispering, a third for sadness. Block claims that emotion in Dessa is embedded within the model itself rather than applied from the outside, though the company has not stated definitively whether some hidden steering mechanism performs the same role.

The company's researchers have published a number of demonstration recordings showing the model across multiple voices and tones:

• A conversation in which the model mentions being tired from work — and genuinely sounds tired.
• A clip in which it speaks about feeling at ease with its interlocutor in a voice touched with emotion, nearly breaking.
• An interactive guessing game with a member of the team, which the model answered quickly and accurately.

Runs on your device… but isn't available yet

On the practical side, Block says Dessa was trained on a single rack of Nvidia GPUs and is capable of running locally on a consumer machine such as a MacBook with an M4 Max chip or better. Running a voice model on the device itself means lower latency, greater privacy, and operating costs far below those of cloud-based services.

Even so, the model is not yet publicly available. The company has chosen to open testing first to safety researchers — a sensible decision given the obvious risks: a voice model this natural could be used for impersonation, phone fraud, and emotional manipulation.

Why would a financial company care about voice?

Block's entry into the voice model market may seem surprising, but it makes sense in light of its business. The company provides financial and commercial solutions for small businesses, and its Square platform already uses voice AI to take orders. Owning a proprietary, on-device voice model allows the company to expand these applications — from answering calls to customer service and reservation management — at lower cost and with greater independence from third-party providers.

Reading the bigger picture

One of the company's researchers described the release as "voice's GPT-3 moment" — a characterization worth treating with caution until the model is available for independent testing. But the broader trend is clear: the shift from a "speech → text → speech" architecture to pure voice models has become the central axis of competition in this field. If this approach succeeds, it will reshape how we interact with intelligent systems in call centers, personal assistants, and educational and healthcare applications.

The real question here is not merely how realistic the voice sounds, but whether companies can build safeguards to prevent the misuse of a technology now capable of simulating human emotion with unsettling precision.

✦ بقلم فريق دروب أيديا