Week ending October 10: Ash finds his voice

This roundup covers October 5–10, 2026. Recon Chinchilla gained a public home, Ash made it through bounded spoken conversations, and we spent time learning the hands-on skills the physical build will need. We also recorded the next piece of Ash’s world: a little mail droid carrying letters from Bishop. Some work reached a working checkpoint; some exposed a problem; some is still a design on paper.

A home for the project. ReconChinchilla.com launched with daily logs, weekly posts, milestones and stories. We brought the earlier engineering history into a readable public journal while keeping the private Project Book and raw evidence separate. The Command Center opened a window into Ash’s animated world, with night vision, character visits and outdoor destinations. It is a visitor experience, not a live connection to the Raspberry Pi or its conversation system.

Saturday’s Command Center update also finished CODD’s coffee routine. CODD—Coffee On Demand Droid—now stops clear of Ash’s foot and raises the tray within reach. Ash picks up the cup with his nearest paw, takes a sip, smiles and returns it before CODD lowers the tray and leaves. The first versions exposed foot overlap and an unnatural sleeve fold; the revised arm movement was reviewed and the final animation approved. The complete coffee-break update is linked below.

Soldering practice belongs in the story too. We spent time with the soldering equipment and snips, and discussed wire-to-pad work as a next practice direction. The difficult part was getting a consistent result: the solder seemed slow to melt, then there was too much of it. That was frustrating, and it remains a skill to develop. Completing a practice session is progress, but it is not proof that a joint is ready for service or that an RC-01 wiring assembly has passed inspection.

After that practice, we cleaned the bench and returned to programming. The early voice-command work recognized a spoken wake-status question and produced “I’m awake and ready.” It also exposed ordinary integration problems: input handling, a quit interaction that needed correction, and phrases the first command vocabulary did not understand. Expanding and checking that vocabulary helped the simple command path before the later conversational work.

Getting the bench usable again took its own effort. Display trouble included static, a blank screen and missing touch input. Touch returned after moving the USB connection to a different monitor port, and display operation recovered. System updates completed, followed by a reported working reboot. Ethernet removed the terminal lag experienced during setup. The monitor’s speaker output remained unresolved, so we kept that limitation visible and used Windows for speech playback.

Ash’s largest software step was a complete local conversation loop. A USB microphone captured speech, Whisper transcribed it, a local Qwen model generated a reply, and Piper produced the voice. The Windows launcher downloaded and played each reply before beginning the next turn. Three-turn sessions demonstrated recall of information supplied earlier in that session. A new launch starts fresh; persistent personal memory is still ahead.

We compared smaller and larger language models and kept the faster smaller model for the current development loop. The model runs on the Pi CPU, not the Hailo accelerator. Listening and speaking happen sequentially, and the conversation interface does not have live file or sensor access. These limits matter: a working spoken exchange is a meaningful checkpoint, but it does not make Ash an always-on, fully commissioned companion.

PMA needed testing as carefully as the audio. Kindness, patience and encouragement are central to Ash’s character, but a personality experiment combined with a repetitive transcript led to repeated, unhelpful replies. We restored the earlier conversational prompt, disabled the recognition hint and added a guard against extreme repeated-word transcripts before they entered the model and conversation history. A fresh three-turn session completed without that loop. The recovery was useful; factual accuracy, invented experiences and natural tone remained open.

A shorter PMA candidate was then tested in text conversations and left uninstalled because it still brought unnecessary reassurance into ordinary exchanges. The lesson was to judge the actual reply, not just a successful program exit or a playable voice file. Support should respond to what someone says, without forcing cheerfulness or inventing a personal experience.

We also gave Bishop a way to stay in touch. The DDS-01 design describes a blue mailbox on treads with expressive eyes, carrying cheerful letters from Ash’s friend in the field. Bishop writes, DDS delivers and Ash chooses how to respond. The plan calls for a 25% delivery chance on Monday, another on Wednesday if needed, and a guaranteed Friday slot if no letter has arrived. Saturday and Sunday each independently offer a 50% opportunity for Ash to initiate a reply while DDS is off regular duty.

The letters will occasionally ask about CODD’s coffee service or the Unknown Engineer. A latest-location-wins notice lets Bishop ask about Ash’s most recent trip; with no new visit, he can gently encourage a break. That design is now recorded in the Project Book and public journal, including event logging and planned tests. DDS implementation, animation, integration and acceptance testing are not done.

By Saturday evening, the voice work had moved beyond that early baseline. The updates below record where we finished the week; the earlier steps remain part of the story.

Ash found his voice. After listening to several options and adjusting the pace, we settled on Michael through local Kokoro synthesis at speed 1.15. The goal was calm, smooth and conversational, with enough pace to flow naturally. The final feedback was simple: “the voice is perfect!” That settles the voice selection for now. It does not mean every answer is ready.

Keeping the mobile robot in view. The Raspberry Pi remains the eventual home of an independent mobile Ash. We used both the Pi and the PC to compare response quality and speed. The PC gives us more room to experiment, while the Pi tells us what the future robot can actually handle. Local voice generation avoids adding another voice subscription. The public Command Center remains separate from this bench conversation system.

Better answers take more than a better voice. We developed limited routes for sourced weather and Eagles questions, selected facts supplied in the conversation, and a few reviewed science answers. We also worked on session handling and recovery when a response service failed. Broader model trials still invented memories or activities and sometimes confused a reported event with an observed one. Automated reviewers missed some of those errors too. General conversational reliability remains unfinished.

The live test showed both progress and a problem. A weather question reached the sourced forecast path and produced a supported reply, but the full trip from listening to an audio file took about 22 seconds. An Eagles question was misrecognized and received an unhelpful answer. A throat-clearing annotation was also treated as conversation. We added a clarification response for the observed ambiguous Eagles phrase and a guard for annotation-only non-speech. Controlled checks passed; another live Eagles exchange is still ahead.

We finished with a real microphone comparison. For the supplied sentence “Ash is a recon chinchilla,” the current recognizer matched the words in about 0.94 seconds. The larger recognizer took about 2.24 seconds and repeated the sentence relative to the supplied reference. Those timings include loading the recognition model and are not full conversation timings. One short sample is not a general accuracy result, but it gave us no reason to replace the current recognizer.

Where we stop for the evening. The approved voice stays as it is. Next comes a complete Eagles exchange through the normal listening path, then fresh conversations that test corrections, uncertainty and honest limits alongside natural PMA. We still need to improve Pi response time, continue soldering practice and inspect the results, and resolve the outstanding speaker-output work. CODD’s website animation is complete; DDS and Bishop’s correspondence remain a documented design awaiting implementation. The Project Book now carries the evening closeout and the ordered plan for resuming.