Key Points
- Block unveiled Dessa, a voice model running locally and producing laughs, pauses and crying sounds.
- The company said Dessa is trained entirely on human speech, using a single NVIDIA GPU rack.
- Block argues text-driven pipelines strip emotional information that voice carries, worsening user experience.
The latest:
A voice model that laughs, pauses and cries has come out of Jack Dorsey’s Block, which said the system runs locally and was trained fully on human speech using a single NVIDIA GPU rack. Called Dessa, it emerged from the company’s research lab. Block said emotion is built into the model rather than layered on afterward.
Details:
- The model: Block described Dessa as a full-duplex system, meaning it can talk and listen simultaneously. That places it in the same category as the GPT-Live model OpenAI released earlier this year, which also handles both directions of a conversation at once rather than alternating between speaking and listening turns.
- The training: The company said Dessa was trained fully on human speech and on a single NVIDIA GPU rack. Block did not disclose the size of the dataset, the length of the training run, or the specific hardware configuration behind the rack.
- No language model: Speech models are typically paired with a large language model for knowledge retrieval. Block said Dessa breaks from that structure, and that whatever the model knows comes from the conversational data it was trained on rather than from an external text-based system sitting behind it.
- The critique: Block said existing voice systems sound unnatural and annoying because text sits in the pipeline driving speech generation. In a blog post, the company wrote that current voice AI “turns your voice into text and throws away an incredible amount of information text alone can’t carry,” calling the result worse and potentially dangerous.
- The competitors: Block named the approach it is rejecting. Models including ElevenLabs v4 and Fish Audio Drama 3 use voice tags to drive emotion in speech generation, an explicit markup layer. Block claims emotion is instead built into Dessa, though the company has not said whether tags of some kind steer that output.
- The demos: Company researchers posted multiple recorded conversations with Dessa showing the agent using different voices, cadences and emotional registers. In one exchange, the bot said it was tired from work and its delivery matched, sounding tired at that point in the conversation.
- Local operation: Block said the model runs locally, meaning inference happens on the user’s own hardware rather than on a remote server. The company did not specify minimum device requirements, availability, pricing, or whether Dessa will be released to developers or kept inside the research lab.
- Also: The release positions Block against ElevenLabs and OpenAI in conversational audio.
Background:
Block is the payments company led by Jack Dorsey, the Twitter co-founder. Dessa comes out of the company’s research lab rather than its core commercial products, and Block did not tie the model to any of its existing payment or financial services.
Between the lines:
Block’s pitch rests on removing text from the pipeline, which is also what makes its emotion claim hard to assess from outside. The company says emotion is built in rather than tag-driven, but it has not ruled out tags entirely. The posted demos show emotional range; they do not show how it is produced.
What’s next
Watch for Block to publish technical details on Dessa’s architecture and training data, any developer access or licensing terms, and whether the company clarifies how emotional output is steered without text tags.
