A draft transcript can contain a fluent sentence that you cannot find in the recording. Natural-sounding text is not proof of spoken words. Use the original audio as the reference before quoting anyone.
Why listening matters
The Whisper model card documents that predictions may include text absent from the audio. This is not a claim that every transcript is wrong, or a measured error rate for FileMyLove. It is a reason to verify quotations.
We recommend paying particular attention to pauses, noise, very quiet speech and unusual repetition. Text outside those areas still needs review.
What the silence check actually does
FileMyLove's audio-to-text tool checks average energy across the entire recording to skip very quiet files. It does not classify every segment as speech or silence. A recording containing both conversation and pauses can still proceed to the model.
There is no automatic sentence-by-sentence verification or guaranteed removal of invented text. The result is editable, but listen before correcting it.
Review without filling gaps by guesswork
- Keep the original recording and a copy of the draft.
- Listen before and after a questionable sentence for context.
- If those words are absent, remove them from the reviewed copy.
- If speech is unclear, mark it as
[unclear]in the working copy rather than guessing. - Recheck names, numbers and negations before quoting the result.
This is our editorial review method, not an automated product feature. Keep unreviewed drafts separate from text ready for use.
Retrying is not verification
You can use trim audio to retry a shorter passage with context, while retaining the full original. Getting the same sentence again does not establish that it was spoken; listen back.
See the Thai interview transcription guide for TXT saving and limits. Audio is processed on your device without uploading it.
