I synthesized twelve seconds of English speech, fed it to a brand-new subtitle workflow, and told it explicitly that the audio was English. The transcript came back word for word correct, three cues, zero mangled words. Then I opened the result JSON to log the run time and found this sitting right next to a perfectly accurate transcript: "lang": "ko".
My first guess was a real bug, the kind where a hardcoded default silently overrides what you asked for. So instead of guessing I went and read the published source. Turns out that particular lang field doesn't describe the audio's language at all. It's driven by a completely separate option, output_lang, that only controls whether the generated review spreadsheet and readme use Korean or English column labels. I'd set language (the transcription option) and never touched output_lang (the document-labels option), so it quietly defaulted to Korean, and the field just happens to share a name with the thing I actually cared about.
Not a transcription bug. The actual speech-to-text handling did exactly what I asked. But a field named lang, sitting inside a transcription result, reporting the spreadsheet's label language instead of the detected or requested speech language, is exactly the kind of thing that would make someone building an automated pipeline log a false alarm.
Zero network requests during the run or the download, checked with a request listener the whole time.
Originally published at toolsthatdontupload.com/blog, where I write up things I actually test on our own free browser tools.
Top comments (0)