Start with French speech, not the interface language
Open French transcription, choose the local recording and select French if you know that is the spoken language. The interface may remain English. Transcription keeps French speech in French; it does not translate it. Auto Detect samples the beginning and chooses one language, so a long introduction in another language can mislead it.
Use the original audible source. Converting a compressed MP3 to WAV cannot reconstruct missing consonants. For an interview, overlapping voices and distant microphones need more listening, whatever the file extension. The interview workflow explains the limits on speaker attribution.
Check the words that alter meaning first
Listen to names, amounts, dates and negatives before polishing punctuation. “Il a quinze dossiers” and “Il a cinquante dossiers” describe different quantities. A name such as Maya is not wrong merely because a dictionary suggests a more common word. Compare the recording, your source notes and the intended spelling.
Homophones are another reason to listen in context. “Ces”, “ses”, “c’est” and “s’est” can sound alike, while their grammatical roles differ. A plausible sentence is not evidence that recognition was accurate. Mark an uncertain passage in your working copy instead of confidently inventing a missing word.
Separate spoken French from writing corrections
The recording might contain “je sais pas” rather than “je ne sais pas”. Adding “ne” may suit a polished document, but it changes a verbatim quotation. Decide whether you need a faithful interview record or edited prose before accepting a suggestion. The same applies to hesitations, repetitions and informal expressions.
Check & Improve uses the local French checker, Grammalecte, alongside deterministic cleanup. It can flag spelling, agreement or punctuation issues, but cannot listen to the source or identify a speaker. Review the original text, proposed replacement and explanation. Accept changes individually when meaning matters; Ignore keeps an intentional form. You can also edit the working transcript manually.
Export and verify the version you actually need
Copy Text and TXT use the selected readable transcript version. SRT, VTT and JSON retain the original timestamped recognition segments. Correcting an agreement in the document does not update a timed cue. Reset to Original discards readable-text changes; save your reviewed copy first if you may need it later.
For captions, download the timed file and check it against the video in a suitable editor. Verify names, line breaks and timing, especially after cutting the video. The SRT and VTT comparison helps choose the destination format. For broader source-quality checks, use the transcript accuracy checklist.