Block, the financial technology company founded by Jack Dorsey, has introduced a new artificial intelligence voice model named Dessa. The system is designed to run locally on consumer-grade hardware, such as a MacBook with an M4 Max chip, rather than relying on massive cloud data centers. According to the company, Dessa is a full-duplex model, meaning it can listen and speak simultaneously, similar to a natural human conversation.
Training on Speech, Not Text
The primary distinction of Dessa lies in its training methodology. Block stated that the model was trained exclusively on human speech data. This approach contrasts with many existing voice AI systems that convert spoken words into text, process that text through a large language model, and then convert the response back into speech. Block argues that this text-based pipeline discards critical information about tone, pacing, and emotional nuance.
In a blog post, the company noted that current voice AI interactions can feel unnatural because they strip away the emotional context of a conversation. Block claims Dessa captures these elements natively. The model produces sounds such as laughs, pauses, and even cries, which are embedded directly into the speech generation process rather than being added via external tags or commands.
Researchers at Block demonstrated the model’s capabilities through several examples. In one instance, the AI agent described feeling tired from work and adopted a corresponding tired tone. In another, the system expressed emotion while discussing a personal topic, sounding almost like it was breaking down. These examples were shared to illustrate the model’s ability to convey nuance without explicit programming for each emotional state.
Local Processing and Safety Testing
Block emphasized that Dessa can operate on a single NVIDIA GPU. This local processing capability allows the model to function without a constant connection to the internet. The company believes this architecture offers significant advantages for privacy and cost, particularly for small businesses that use Block’s Square payment systems. Voice AI is already used in Square to handle ordering, but a local model could expand these applications to more consumer and business experiences.
Despite the technical achievements, Block has not made Dessa available to the general public. The company is currently inviting safety researchers to test the model. This phase is intended to identify potential risks, such as scenarios where a voice agent might deceive users or exhibit unintended behaviors. The company described the release as a significant milestone for speech technology, comparing it to a pivotal moment in the history of large language models.
The move comes as the broader tech industry continues to refine voice interaction. Competitors like OpenAI have released models such as GPT-Live that also support full-duplex conversation. However, Block’s focus on local execution and emotion-driven speech sets its approach apart from cloud-heavy alternatives. The company did not disclose specific pricing or a public launch date for Dessa.
Source: TechCrunch
