Observed (real capture data, 2026-08-23)
Dictating an email address of the form <name>@<provider>.com produced, across four attempts:
| spoken |
transcribed |
| the address |
<Name split into words> at <provider>.com. |
| (retried) |
<digits> at <provider>.com. |
| (retried) |
<First name> at <provider>.com. |
Four attempts, four different wrong answers. @ is never produced; "at" and "dot" stay as words, and the local part is heard as ordinary words or numbers.
Why the polisher can't fix it
Same class as "TimbreKit" → "spell kid": the transcriber lands somewhere the polisher can't recover from. A local part heard as a plausible English phrase carries no signal that it was ever an address.
Candidate approaches
- Deterministic address assembly (ADR-0005 principle): when a transcript matches
<words> at <words> dot <tld>, rebuild it as an address. Mechanical, has a right answer, testable. Risk: "meet me at gmail dot com" is not an address — needs a tight pattern, probably requiring a known TLD and no intervening punctuation.
- Vocabulary at the transcriber via
SFCustomLanguageModelData — teaches it the local part as a token. Helps the name, not the @.
- Prompt rule — rejected on precedent: GDR-0003 and ADR-0005 both found the model unreliable at mechanical transformations.
(1) plus (2) is likely the answer. Needs a corpus of real dictated addresses before tuning.
Timing data from the same capture
Cold Bluetooth mic wake measured at 1890 ms vs 291 ms warm — closes the measurement half of issue #4.
Replaces #13, which quoted a real address and was blanked when the repository went public.
Observed (real capture data, 2026-08-23)
Dictating an email address of the form
<name>@<provider>.comproduced, across four attempts:<Name split into words> at <provider>.com.<digits> at <provider>.com.<First name> at <provider>.com.Four attempts, four different wrong answers.
@is never produced; "at" and "dot" stay as words, and the local part is heard as ordinary words or numbers.Why the polisher can't fix it
Same class as "TimbreKit" → "spell kid": the transcriber lands somewhere the polisher can't recover from. A local part heard as a plausible English phrase carries no signal that it was ever an address.
Candidate approaches
<words> at <words> dot <tld>, rebuild it as an address. Mechanical, has a right answer, testable. Risk: "meet me at gmail dot com" is not an address — needs a tight pattern, probably requiring a known TLD and no intervening punctuation.SFCustomLanguageModelData— teaches it the local part as a token. Helps the name, not the@.(1) plus (2) is likely the answer. Needs a corpus of real dictated addresses before tuning.
Timing data from the same capture
Cold Bluetooth mic wake measured at 1890 ms vs 291 ms warm — closes the measurement half of issue #4.
Replaces #13, which quoted a real address and was blanked when the repository went public.