Block Unveils Dessa: AI Voice Model Trained Entirely on Human Speech Captures Emotional Nuance

Block Introduces Speech-to-Speech AI That Processes Pure Audio

Jack Dorsey's financial technology company Block has unveiled Dessa, an AI voice model trained entirely on human speech without relying on text or pretrained models1

1

5

5

. Unlike conventional voice AI systems that convert speech into text and back, Dessa operates as a speech-to-speech model, processing audio directly to preserve emotional cues and variations in tone that typically get lost during text conversion2

2

5

5

. The company trained Dessa on a single NVIDIA GPU, specifically an NVIDIA GB200 NVL72, despite the complexity of training on speech data which contains significantly more information than written text5

5

.

Full-Duplex Architecture Enables Natural Conversation Flow

Dessa is a full-duplex voice AI model, meaning it can listen and speak simultaneously, mirroring how humans naturally converse1

1

4

4

. This architecture prevents the abrupt silences typical of conventional voice assistants that stop speaking the moment a user interrupts3

3

. Block emphasized that current voice AI "turns your voice into text and throws away an incredible amount of information text alone can't carry," arguing that this limitation creates worse experiences and potentially dangerous situations1

1

. The model can independently determine what to say, when to respond, and how to express itself, including choosing its voice and conversational style5

5

.

Emotional Nuance in AI Creates Uncanny Human-Like Interactions

Demonstration clips show Dessa expressing a wide range of emotional cues including laughter, pauses, frustration, and what sounds like teary gratitude1

1

2

2

. In one example, when told about a surprise trip to Paris, Dessa responded with audible gasps and joyful enthusiasm that verged on tears4

4

. Another clip featured Dessa speaking with the unmistakable cadence and intonation of an African American man, discussing being tired from work and a newborn baby named Darius Johnson3

3

. The model can also detect sarcasm, a sophisticated understanding of tone that has historically stumped AI systems4

4

. Block claims that unlike models such as ElevenLabs v4 or Fish Audio Drama 3 that use voice tags to drive emotion, Dessa has emotion built directly into the model, though it remains unclear if emotion is being steered by tags1

1

.

Local Deployment Raises Business Applications and Safety Concerns

Dessa can run locally on consumer-grade computers such as a MacBook Pro M4 Max or better, enabling on-device deployment without cloud dependencies1

1

2

2

. For Block, which operates Square point-of-sale systems already using voice AI for ordering, a proprietary voice model could power significantly more business and consumer experiences at lower cost1

1

. However, Block is not offering public access yet and has instead enlisted a small group of AI safety researchers to test the model for risks the company hasn't anticipated1

1

5

5

. The primary concern centers on impersonation risk, where a voice agent could potentially trick people on the other end of conversations1

1

. Block stated that AI "keeps gaining power and independence, and it's still nearly deaf to the emotion in our voices," positioning Dessa as addressing the safety problem of AI being unable to recognize urgency, fear, or pain in voice patterns4

4

. The Dessa team cited three motivations for development: curiosity, better understanding between humans and AI, and safety2

2

. As human-AI interaction evolves toward more natural speech interactions, the challenge lies in ensuring these increasingly convincing models don't enable harmful deception while improving genuine communication.