Weekend Engineering Review October 9 to 11 2026
Voice selection, bounded demonstrations, failures and open validation gates.
Published 2026-10-12 · Dated evidence, not a live status feed
Weekend Engineering Review October 9 to 11 2026
We made substantial progress toward giving Ash a voice, and we ended the weekend by stopping to reconcile our records with the system.
The accepted milestone is specific: Kokoro's Michael voice at speed 1.15 delivers the calm, conversational sound we wanted. That acceptance does not close the work on recognition, truthfulness or natural conversation. Keeping those decisions separate makes the next tests meaningful.
Mission objective
Develop a bounded speech pipeline, assess voice and response quality, preserve a recoverable baseline, and record the difference between a working component and an accepted integrated system.
System overview
The demonstrated recovery route captures speech through a USB microphone, transcribes it with native Whisper, generates a local model reply, synthesizes a WAV file and plays that reply on Windows. The listener stops before reply generation. Sessions are bounded to three turns and reset on a fresh launch.
The earlier v12 recovery path uses Qwen 1.7B and Piper. The later isolated development candidate uses Qwen 4B and Kokoro Michael 1.15. These paths coexist. Selecting the everyday launcher remains an open decision. The owned PC is also a comparison platform; its faster results do not qualify standalone Pi performance.
The Hailo accelerator remains part of perception. The local conversation model runs on the Pi CPU; we did not establish Hailo language-model acceleration.
Architecture
USB microphone → native Whisper transcript → bounded conversation session → local language model or supported source route → speech synthesis → validated WAV transfer → Windows playback.
Perception → state → behavior remains a separate managed service chain. Coexistence with the voice candidate does not establish integration.
What the evidence supports
See the retained status register.
| Workstream | Result | Limit |
|---|---|---|
| Display and touch | User confirmed restored operation after recovery | One root cause for every symptom was not established |
| Network | Wired 20-packet sample had no loss and averaged 45 ms; user reported lag resolved | Not an endurance qualification |
| Recovery conversation | Three-turn microphone-to-Windows-playback sessions demonstrated | Not always-on; no persistent memory |
| Voice selection | Michael 1.15 accepted | Recognition and answer quality remain open |
| Candidate tests | Last full Pi candidate suite recorded 139 passing tests | Later focused checks are separate; no aggregated new full-suite count |
| Writer correction | Installed source/binary identities verified; eleven isolated checks passed | Installed listener not executed in that deployment; live integration open |
| Managed services | Perception, state and behavior active in October 11 snapshots | Not proof of continuous health to the cutoff or voice readiness |
| Character development | Eleanor and Dr. Jenny reference sheets available | Reference sheets are not finished rigs or deployed animation |
| CODD coffee routine | Public website routine completed and approved in prior record | Virtual animation, not physical actuation |
| DDS and seasonal Bishop concept | Design recorded | Implementation and tests not done |
Failures retained
An earlier recognition failure produced an extreme repeated transcript. The subsequent conversation carried that bad input forward. Recovery v12 restored the prior prompt, disabled its vocabulary hint and rejected a narrowly defined extreme-repetition pattern before inference and history updates. A successful subsequent session supports that recovery scope, not general recognition acceptance.
A later live weather question reached a supported source route and Windows playback. The recorded time to the audio file was about 21.70 seconds, including listening, transcription, retrieval and synthesis. This is one complete observed trial, not a latency guarantee.
The Eagles question was transcribed ambiguously and missed its route. Throat-clearing was incorrectly treated as conversation. Focused repairs now request clarification for the observed ambiguous phrase and keep annotation-only non-speech out of conversation history. Three non-speech tests and fifteen Eagles tests passed, followed by controlled text replay. A new live microphone Eagles test remains unrun.
The development work also exposed invented memories, unsupported activity claims and tentative plans treated as facts. Separate reviewers and a claim-support classifier sometimes approved unsupported claims. Those checks cannot serve as an acceptance gate.
Validation checkpoints
With the owner’s freeze release recorded, the next bounded test must establish the exact launcher, source/binary identities, real capture, persistence, response generation, WAV validation and actual playback for the same turn. Recognition, factual support, voice quality and latency must each receive their own result.
Independent speaker operation, array qualification, live failure cleanup, endurance and recovery remain separate gates. No wake word, barge-in, always-on conversation or persistent memory has been accepted.
Next action
The owner released the runtime change freeze on October 12, 2026. That release permits resumed development; it does not turn open validation into acceptance. This website update performs no device or speech work. Next: select a bounded regression objective and retain its evidence.