Skip to content

Cleaning up a transcript with the LLM

How to ask the language model for a tidier copy of a transcript, check every word it took out, and download the result.

People rarely speak in clean sentences. A transcript that is faithful to the recording is full of "I mean", "you know" and false starts, which makes it hard to quote or share. Voxint's filler removal already drops fixed words such as um and uh. The LLM clean-up goes a step further: the language model (LLM) you set up reads each line and suggests which words are filler in that context, so it can catch phrases a fixed list cannot.

The clean-up is a separate copy. Your reviewed transcript, its normal downloads and its subtitles stay exactly as they are, and the clean-up appears only where you ask for it.

What the model is allowed to do

The model can only suggest words to remove. Voxint checks every suggestion against the original line before using it:

  • Removing words is the only change allowed. If the model rewords a line, adds a word or swaps one word for another, Voxint ignores its suggestion for that line and keeps the line as it was.
  • Some words are never removed: numbers (written as digits or as words), negations such as "not", "never" and "don't", and words that look like names (a capital letter in the middle of a sentence). A suggestion that removes one of these is ignored for the whole line.
  • A whole line is never removed. A suggestion that would empty a line is ignored.
  • The cleaned text is built from your transcript's own words. The model never contributes a character, so the copy cannot contain a word that was not in the transcript.

The clean-up is the model's reading of what counts as filler. Check the struck-through words before you use the copy. It does not cut the audio, and leaving a word out is not redaction: the full transcript is still stored.

Before you start

  • Turn on the LLM. The clean-up uses the same LLM as transcript enhancement. Open Settings, find the LLM section, and make sure LLM transcript enhancement is on. See Settings & troubleshooting if you have not set up the LLM yet.
  • English recordings only. If Voxint detected another language in the recording, the clean-up is not offered. If it could not detect the language, you can still run it, and the page notes that the language was not verified as English.
  • Optional: your filler words. The words in Settings → Filler words are passed to the model as examples of what to look for. See Choose your filler words.

Make a clean-up

  1. Open the recording in the editor (from Media, click the recording).
  2. Click Download transcript, expand Other text variants, and click LLM clean-up. The reading view has the same LLM clean-up link at the top.
  3. On the clean-up page, click Generate clean-up. The page shows Cleaning up with the job's progress and a Cancel button. A long recording can take several minutes, depending on your LLM.

When it finishes, the page reloads with the result. (With JavaScript turned off, reload the page yourself to see when it is done.)

Check what was removed

The Changes section shows every line of the transcript with the removed words struck through, so you can read the original and the cleaned version in one pass. A line where Voxint ignored the model's suggestion is shown whole, with a short note saying why, for example "(kept: reworded, not only shortened)".

The Summary above it gives the totals: how many lines changed, how many words were removed, and how many suggestions were not used, by reason. It also records which model made the clean-up and when.

Download the cleaned copy

The Download section on the clean-up page lists the cleaned copy in every transcript format: plain text and Markdown (each as a reading copy or with times), .srt and .vtt subtitles, and .json. Subtitles keep the original timing: each caption shows the cleaned words for the same stretch of audio.

The cleaned copy is not combined with the reading view's Remove filler words or Remove repeated words options, or with a translation. It is its own version of the text, and the words you marked to keep or leave out in the review editor do not apply to it.

If you edit the transcript afterwards

A clean-up is a snapshot of the transcript at the moment it was made. If you correct a line or split a segment afterwards, the clean-up no longer matches, and the page says so:

  • An Out of date notice appears at the top.
  • The download links are removed, and a direct download is refused rather than served stale.
  • The Changes section still shows the old clean-up against the transcript as it was then, so you can see what it did.

Click Generate again to clean up the current transcript. The new clean-up replaces the old one. Renaming a speaker, changing your filler words, or marking words in the review editor does not make a clean-up out of date; only changes to the transcript text do.

Next steps