Block Unveils Dessa: AI Model Trained Entirely on Human Speech Without Text

Block Introduces Speech-Native AI Model

Jack Dorsey's financial technology company Block announced Dessa, an AI model that represents a fundamental shift in how voice assistants learn language1

1

2

2

. Unlike conventional systems that convert speech to text and back, Dessa was trained from scratch exclusively on human conversation data without using any text or pretrained models. The speech-to-speech model processes audio directly, preserving the emotional cues and tonal variations that typically disappear during text conversion.

Training Challenges and Technical Achievement

Block faced significant obstacles in developing Dessa since spoken words contain substantially more data than written text and vary with each utterance2

2

. Despite these complexities, the company successfully trained the AI model using only one NVIDIA GB200 NVL72. The team noted that Dessa learned language from hundreds of times fewer words compared to large language models, which require massive text datasets. The model can run locally on a MacBook Pro M4 Max or better hardware, making it accessible for deployment without cloud dependencies1

1

.

Natural Speech Interactions and Emotional Intelligence

Dessa operates as a full-duplex conversational system capable of listening and speaking simultaneously while recognizing and expressing emotions including laughter, sarcasm, frustration, and sadness through natural speech interactions2

2

. The AI model independently determines what to say, when to respond, and how to express itself, including its voice and conversational style. Block shared demonstration videos showing Dessa playing guessing games and engaging in discussions about topics like Christopher Nolan's films, though the conversations occasionally exhibited awkward moments, such as mistaking a speaker for someone named "Ronaldo"1

1

.

Comparative Performance and Real-World Applications

The Dessa website features comparisons between the model and OpenAI's GPT-Live-1 voice model, where Dessa demonstrates more natural conversational flow1

1

. In one example, a user describes planning a surprise trip to Paris, and Dessa responds with what sounds like genuine emotional gratitude. The model reacts naturally to humor by laughing mid-sentence and adjusts its speaking volume when requested, behaviors that emerged without explicit training for these responses2

2

.

Safety Considerations and Future Access

Block acknowledges that training an AI model exclusively on sound represents unexplored territory with potential risks that require external evaluation2

2

. The company is granting a small group of AI safety researchers early access to identify risks that the development team may not have anticipated. The Dessa team cited three primary motivations for creating the model: curiosity, improving understanding between humans and AI, and safety1

1

. This cautious approach to deployment signals awareness of the novel challenges posed by speech-native AI systems that bypass traditional text-based safeguards.