Finding Ash's Voice
Michael 1.15 is the selected development voice; recognition and factual reliability remain open.
Record date Oct 12, 2026 · Published
After successive listening samples and live development work, we selected Michael at speed 1.15. The delivery is calm, clear and conversational. It is a meaningful creative milestone supported by a working speech pipeline.
The engineering challenge extends beyond producing a pleasant voice. Ash must understand what was said, avoid inventing facts, respond naturally and handle uncertainty without becoming repetitive. Our tests exposed problems in each of those areas, and those failures remain in the project record.
We demonstrated bounded spoken sessions through the Pi with playback on Windows. We also tested focused repairs for ambiguous recognition, non-speech input and response integrity. Those results move the project forward, but they do not establish an independent, always-on robot.
The weekend ended with a change freeze after finding a mismatch between the system and its documentation. The owner subsequently released that freeze on October 12. Bringing the record current is part of building Ash responsibly. We want every milestone to have a clear meaning and every next step to start from a configuration we understand.