TranscribeLocally

How to clean up a machine transcript quickly

An automatic transcript is a first draft. The difference between one that took twenty minutes to finish and one that took two hours is almost entirely the order you do things in — most people start at the top and read every word, which is the slowest possible approach.

Here is a faster pass. It assumes roughly fifteen to twenty-five minutes of editing per hour of audio for a clean recording, more for a difficult one.

First, decide what the transcript is for

This determines how much work it needs, and it is worth two seconds of thought because the answers differ enormously.

Most transcripts are the first three, and most people edit them as if they were the fourth.

The order that saves the most time

1. Fix recurring proper nouns first, globally

The largest error category is names — people, companies, places, products. They are also the easiest to fix, because they repeat: one wrong name may appear thirty times.

Correct the first occurrence, then use the search box to find every other instance. Two minutes here removes more errors than twenty minutes of reading linearly, and it removes the errors readers actually notice.

Do the same for specialist vocabulary: the technical term your field uses that a general model has never encountered.

2. Delete the invented text

Look at the very beginning and the very end, and at any point where music or a long silence sits in the recording. Whisper is trained to produce speech, so given no speech it sometimes produces a plausible sentence anyway — often a stock phrase or a repeated line.

Runs of identical repeated segments are stripped automatically here. What is left tends to sit at obvious boundaries: a sign-off phrase over the outro music, a thank-you over thirty seconds of silence. Delete them.

3. Label the speakers

Do this before any detailed reading, because it changes how the transcript reads and it is quick. Click a line, name the speaker, and the label applies to that line and every one after it until the next label. An interview needs one click per turn change; a two-host podcast needs a handful.

This is manual because it is the only way to be right. Automatic speaker identification is a separate model, and where it fails — similar voices, crosstalk — it produces confident, wrong attributions. In a transcript with quotes in it, that is worse than no labels.

4. Check the numbers

Every dose, date, price, version, case reference and phone number. Numbers are where automatic transcription errs in ways that read perfectly fluently, and they are usually the part someone will act on.

Click the line and listen. Do not assume.

5. Now read it — with the audio, not without

Play from the top and read along, clicking any line to jump. Reading against the audio is far faster than reading alone, because you catch the errors that look correct. A sentence that says something plausible but wrong is invisible on the page and obvious in half a second of playback.

Keyboard shortcuts make this bearable: space or K to play and pause, J and L to skip back and forward ten seconds, / to jump to search.

What to leave alone

Over-editing is the other way to waste an afternoon.

Filler words. “Um”, “you know”, “like”. Whether to remove them is a decision about what the transcript is for, not a correction. Verbatim transcripts keep them because how someone spoke is part of the data; readable transcripts remove them. Pick one and be consistent — but do not remove them from a research transcript out of tidiness, because hesitation is often the finding.

Grammar in speech. People restart sentences and leave clauses unfinished. That is what they said. Tidying it silently changes the record.

Punctuation you would have placed differently. It is predicted from the rhythm of the speech and is usually defensible. Only fix it where it changes the meaning.

Paragraph breaks. They are grouped at natural pauses. Fine as they are unless you are publishing.

Choosing the right export

The format does some of the formatting for you:

All of them use your corrections, so edit first and export once.

If the transcript is bad enough to restart

Sometimes the right answer is not editing. If more than roughly one word in five is wrong, fixing it takes longer than a better approach:

A realistic finishing standard

For most purposes, “every name, number and quoted passage is verified; the rest is readable and substantially correct” is the right place to stop. Chasing a perfect verbatim record of a conversation nobody will read word-for-word is time spent on the wrong thing — and if the record genuinely must be exact and complete, that is a job for a human transcriptionist rather than a longer editing session.