How to edit video by editing text, with cuts that sound clean

What is text-based video editing?

Text-based video editing is editing a video through its transcript: you read the words, delete the ones you do not want, and the editor removes the matching picture and sound. It works because every word in an accurate transcript is tied to the moment it is heard, so a sentence is also a stretch of time.

It suits anything carried by speech: talking heads, interviews, podcasts, tutorials.

The last row matters. A transcript's word times are estimates: cut exactly on them and you hear the tail of a deleted word, or lose the first sound of the next.

Can you edit a video by editing the transcript?

Yes, as long as the editor ties every word to the sound it came from. Sharpy makes the transcript on your Mac with two transcription engines that check each other, in 4.9 minutes per hour of sound on the base Mac mini M4 with 16 GB. Then:

1. Open the words view. The WORDS tab in the left panel lays out the words of every recording in the edit, in the order they are heard. A word the transcribers were unsure of is underlined.
2. Read and listen. Click a word to put the playhead there. As the edit plays, the word being heard is lit.
3. Select and delete. Select words and press Delete. They leave the picture and the sound together, as one step.
4. Read what you cut. Cut words stay in the text, struck through and dim. A long run folds to one mark, such as "⋯ 120 words cut ⋯".
5. Put words back. Right-click and choose Put Back, and the cut fills again from the original media. ⌘Z takes a cut back in one step.

Cuts made anywhere, on the timeline or by the agent, show in the text too. Deleting words moves only the speech, so B-roll and music stay on their words. To find a line, press ⌘F and type it: Sharpy searches what is said in every recording, plus captions, markers and clip names, and the list narrows as you type.

The agent does the same with remove_words, naming words rather than frames.

What does the review list find?

The pass an editor makes over somebody talking, as one list. Each item shows its words in colour inside their sentence, the reason, and where it plays.

Each kind has a chip with its count. Hear plays an item with a second and a half either side; then choose Cut (or Shorten, for a pause) or Keep. Cut all deals with every item of the kind in view as one step. A retaken line shows a numbered chip per attempt: click one to play it and keep it, then choose Cut the others.

The list knows what to leave alone. "Mm-hmm" means yes, so it is not a filler. Words said twice on purpose, "very very" or "no no", are kept. A pause the story wants, for emphasis or a reveal, is not listed, and a long stop is shortened, never closed up. Sharpy reads the sentence before it calls "like" a filler, and with several people talking it finds restarts and retakes per speaker.

On a 21-minute, four-person meeting whose fillers were labelled by hand, 49 of the 56 ums and uhs the list put forward were real ones. The agent works through the same list with review_words. How to remove filler words and dead air covers that clean-up pass in depth.

Where does each cut land, so it sounds clean?

Between words, found from the sound. Sharpy measures where each word really starts and ends and puts every cut between one word's last sound and the next word's first: inside the silence where there is one, at the quietest moment where the words run together. In tests across five recordings, cutting on the transcriber's own word times left a deleted word still audible in a quarter of cuts; placing cuts from the sound brought that to about one in ten.

The pause left behind comes from the speaker. Inside a phrase, none. At a sentence end, the short pause this speaker takes there, measured on the recording being cut. Ask the agent to read the pauses (read_pauses) and it labels each by what it is for, a breath, a stall, a beat before a reveal, and every word cut after that, the agent's and yours, follows the reading. Each cut also gets a crossfade with the room's own sound under it, so a join never drops into dead silence. How Sharpy decides where to cut walks through these decisions on a real take.

How do you fix a word the transcript got wrong?

Select the words, right-click, choose Correct and type what was said. It can be more words or fewer ("gonna" for "going to"), or words nobody heard at all; Correct Everywhere sits beside it. The new words are placed in the sound again in about two seconds, and the correction is kept by where it was said, so it survives a fresh transcription. The agent uses correct_words.

For names that keep coming up, give the agent your own vocabulary with add_to_vocabulary: each term and how it gets misheard, such as "Claude Code" heard as "cloud code". It changes only places that are the term spelled another way or a listed mishearing, so "code" never becomes "Codex", and it fixes the captions, the words view and what the agent reads without hearing the recording again. On a test recording, five terms came out right in all three takes, in 11 changes, none wrong.

What should you check before you export?

• Every join, by ear. In connected speech there is sometimes no clean instant between two words, so play each cut once.
• The underlined words. The transcribers were unsure of them; correct any that are wrong.
• Captions made before a correction. They keep the misheard words; the agent is told which.
• Speed-ramped clips. Their words cannot be placed exactly, so they refuse to be cut. Cut the words first, then ramp.

The next lesson, choosing the best take of every line, picks up where retakes leave off.