2026-08-25
Phare The Gosway
Voice is the whole reason this thing works. One thumb, one press, standing at a letterbox. If capture is any harder than that it stops happening, and a memory nobody feeds is worse than none.
So we tested transcription properly before trusting it. Clean synthetic speech , the easy case, no background noise, no accent.
English. "…the policy number is four zero zero eight, three two five, eight seven seven." Came back as 4008-325-877. Perfect, and it grouped the digits correctly.
German. "…die Stammnummer lautet zwei eins drei Punkt null sechs sechs…" Came back as 418.229.703. Exact, dots and all. The best of the three.
French. "la plaque c'est ZH 318602" came back as CBI 318 602.
The digits survived. The canton prefix didn't, and CBI 318 602 still looks like a licence plate. That's the failure that matters: not obviously broken, just quietly wrong, discovered in two years at a garage counter.
Then the case that actually describes this household
Almost every document here contains two languages. French headings above German tables is normal, not an edge case. So we tested a sentence that switches mid-breath, which is how people here actually talk:
Said: *"le Selbstbehalt est de mille francs… la franchise casco
partielle… le Fahrzeugausweis est dans la voiture"* >
Heard: *"le Selbstbehalte… la franchise casque au partiel… le
phare The Gosway est dans la voiture"*
A vehicle registration document became a lighthouse.
The words destroyed are precisely the low-frequency administrative vocabulary this corpus is made of: the terms you'd need to quote on a form or to an insurer. The matrix language was fine. Code-switching inside one utterance was not.
So the rule for the two of us is one sentence long. One language per voice note. A German term inside a French sentence is exactly where it breaks. Say it entirely in German, type it, or photograph it.
And anything with a code on it gets photographed. Not because photos are better. They have their own failures. But because a mangled code looks exactly like a correct one, and there's no way to tell from the transcript.