Repairing mangled names — and the fix that broke everything

Speech recognition garbles names and figures. My first attempt to auto-repair them “fixed” Whichever into Whisper.

Speech recognition mangles the two things you most need it to get right: names and numbers. A colleague’s name comes out three different ways in one meeting; a figure lands as “fifty” when it was “fifteen.” So I built a repair pass that fixes garbled proper nouns back to their correct spelling. My first version was too clever, and it caused a small disaster.

The idea was to take a word that sounded like a known term and snap it to that term. Run over real transcripts, it cheerfully “repaired” Whichever into Whisper, a company’s web address into just its name, and Outlook’s into Outlook — because a generic list of known words collides with ordinary English, and once you let a fuzzy matcher loose it finds matches everywhere. It was correcting things that were never wrong.

Note that the fix was not a smarter matcher. It was a much smaller dictionary. The repair pass now draws only from words the user was actually told about — the people on the meeting invite, and their own glossary — not a generic jargon list. If the system was never told a name is relevant, it does not touch a word that merely sounds like it. Precision over recall, again: better to leave a mangled name than to “correct” a normal word into a name.

The general lesson is that a tool with the power to change your data needs a very good reason before it uses it. The bug was not that the matching was inaccurate; it was that I gave it authority over too much. Narrow the authority — only repair against terms you were explicitly given — and the same mechanism goes from dangerous to useful.

More soon.