Block announced Dessa, a speech-to-speech AI model trained exclusively on raw human speech rather than text transcripts. By learning directly from audio without converting inputs into written text, the model aims to capture emotional tone, vocal nuance, and conversational timing that traditional voice interfaces flatten or ignore.

According to the Dessa development team, the system can run locally on Apple Silicon hardware, specifically a MacBook Pro M4 Max or higher. Demonstration clips released by the company showcased natural back-and-forth dialogue, vocal expressions of gratitude, and interactive audio games, alongside comparison tests with OpenAI’s GPT-Live-1 voice model.

The team framed the project as both an exploration of speech-first learning and a safety initiative, noting that recognizing human distress or emotion in real time is critical for safe voice interactions. Block plans to grant early access to a select group of AI safety researchers to evaluate the model.

Why it matters

  • Demonstrates native speech-to-speech architectures running locally on high-end consumer hardware like Apple M4 Max chips.

  • Addresses emotional detection in voice interfaces, which is critical for real-time customer support, healthcare, and safety.

  • Bypasses text-based intermediates to reduce latency and improve natural timing in spoken human-AI interactions.

Source: theverge.com