Block has released Dessa, an artificial intelligence model trained exclusively on human audio recordings without any text data. This system processes speech directly to interpret emotions, sarcasm, and pauses in real time, bypassing the standard method of converting sound into text first. The software supports full-duplex communication, allowing it to listen and speak simultaneously while running locally on a MacBook Pro M4 Max. Block developed the tool out of curiosity regarding better human interaction and safety concerns, noting that existing voice models often ignore emotional cues. The company has not made the model publicly available yet. Instead, they are inviting AI security researchers to test the system for unexpected risks via a sign-up form.
This approach addresses a limitation where previous attempts at emotional voice AI focused primarily on sounding natural rather than demonstrating actual intelligence. Projects from Hume AI, Kyutai, and OpenAI have improved intonation and empathy, but reasoning capabilities remain more critical for answer quality than vocal tone. Dessa represents a technical shift toward processing raw audio to capture nuance that text-based models miss. It highlights the distinction between simulating emotion and understanding it within a conversation.
- Trained entirely on human speech with zero text data
- Runs locally on MacBook Pro M4 Max hardware
- Currently restricted to security research testing
