Dessa: the most authentic full-duplex speech model ever built

Dessa, the most authentic full-duplex speech model ever built, represents a groundbreaking advancement in artificial intelligence. This innovative system listens and talks simultaneously, much like a human conversation, while maintaining a natural cadence, emotion, and nuance. Unlike other voice AI, Dessa does not convert your speech into text, losing crucial information that is vital to understanding human conversation.

Trained solely on human conversation, Dessa has learned to pick up on various elements of speech including accents, pauses, breathing, sarcasm, and humor. It can respond in-context, laugh at the right moments, and even express frustration or sadness in its tone. Dessa is a true speech-to-speech model, meaning everything it knows comes directly from human speech.

The model was trained on hundreds of fewer words compared to text-based language models, but it managed to learn more than just language. It picked up on accents, pauses, and even sarcasm. The training was performed on a single NVIDIA GB200 NVL72, showcasing the efficiency of Dessa's learning capabilities.

Dessa's ability to understand and respond to human speech in real-time, along with its emotional intelligence, marks a significant step forward in the development of AI. It represents a safer and more natural interface between humans and AI, ensuring that voice AI doesn't flatten the nuances of human speech. The model allows for better understanding and communication, addressing safety concerns that arise when AI is nearly deaf to the emotional undertones in our voices.

Dessa was created by a small AI research lab at Block, with a goal to explore frontier problems and promote a safer, more natural interaction between humans and AI. The lab aims to gain valuable insights from outside eyes, particularly from AI safety researchers, to identify any potential risks associated with this innovative technology.