Transcript cleanup needs an editing policy, not a global filler-word filter

A text-cleaning step can quietly change what a speaker meant. Consider this invented example:

I sent fifteen—no, fifty invitations. Um, I haven’t received a reply.

Removing an empty “um” may fit a clean-verbatim brief. Deleting the self-correction or “haven’t” changes the statement. Treating every repeated word as noise can also erase emphasis or an answer.

Keep the source and the editing policy

For a transcript-processing workflow, keep the source recording and raw transcript alongside the edited reading copy. Write down what your output may remove, how you mark unclear speech, and whether you preserve false starts. Apply that policy before making blanket substitutions.

A useful review sequence:

1. Compare names, numbers, negation and corrections with the recording.
2. Keep speaker turns and meaningful short replies.
3. Mark uncertainty instead of guessing.
4. Check that punctuation hasn’t created a new meaning.
5. Label a summary as a summary; don’t present it as verbatim speech.

Treat cleanup as an editorial transformation

“Clean verbatim” and “full verbatim” are not perfectly universal specifications. Agree on the editing brief for the intended reader. A readable meeting note and an oral-history transcript can require different treatment of hesitation and non-speech events. A global regex cannot decide what an utterance means in context.

Disclosure: I work on Wordtake. The full guide has paired examples, source references and a reusable brief:
Clean Verbatim vs Full Verbatim: What Should You Keep?

The examples are invented. This is an editing workflow, not a measured accuracy benchmark. Prepared with AI assistance.