Text-Based Video Editing
Text-based video editing is a way of editing video through a time-aligned transcript instead of a timeline. Every spoken word carries a start and end time, so selecting, deleting or rearranging text cuts or moves the matching audio and video. It makes dialogue editing fast, but a transcript cannot describe silent shots, framing, color, music or effects.
Three Things “Edit Video with Words” Can Mean
1. Edit the transcript
Select existing words as if the recording were a document. Delete or move them and the media follows. This page is about that workflow.
2. Describe the edit
Type an outcome, like “keep the best answer and add captions,” and an AI editing agent performs it against the transcript and frames.
3. Generate from text
A text-to-video model synthesizes new frames from a prompt. That is generation, not editing a recording.
How It Works
- Transcribe with per-word timing. Speech recognition produces word-level timestamps: a start and end for every word, typically on a 20 ms grid.
- Bind text to time. Each word record holds its position in the text and its span in the source, so a selection resolves to a time range and the playhead resolves to a word.
- Resolve the deletion. Deleted words drop out of a keep list of surviving source ranges, with edges snapped to word boundaries so no word is cut mid-syllable.
- Conform the picture. At export, the ranges are cut from the original file and rounded to whole frames. Nothing is re-encoded until then, which is why deleting a paragraph is instant.
words: [
{ "text": "we", "start": 12.437, "end": 12.556 },
{ "text": "shipped", "start": 12.556, "end": 12.921 },
{ "text": "it", "start": 12.921, "end": 13.024 },
{ "text": "on", "start": 13.024, "end": 13.160 }, <- 352 ms pause follows
{ "text": "Friday", "start": 13.512, "end": 14.005 }
]Inside a phrase, each word ends exactly where the next begins: continuous speech has no gap to measure. The 352 ms before “Friday” belongs to neither word. It is a breath, and it is where a cut can hide.
Which Editors Support It
| Editor | Transcript workflow |
|---|---|
| Descript | Delete, cut, copy and paste script text; the linked media follows and can be restored. |
| Premiere Pro | Select, cut and rearrange transcript text to trim and place clips; filter fillers and pauses. |
| DaVinci Resolve Studio | Transcript-driven rough cuts and timeline editing (a Studio feature). |
| VEED | Cut, restore, remove pauses and filler words from the Edit by Script panel. |
| CapCut | A transcript panel for removing filler words and speech gaps before returning to the timeline. |
Behavior checked against each vendor's documentation on August 24, 2026 (sources below). The idea is older than any of them: research systems such as Carnegie Mellon's Silver (2002) and QuickCut (UIST 2016) linked editable transcripts to video years before Descript made the document the main editing surface.
Where It Runs Out
- It only sees speech. Silent b-roll, music, a reaction shot or an unnarrated demo produce no text to select.
- It says what, never how. “Punch in here,” “grade this warmer” or “drop the music under this line” are not deletions of text.
- One caret cuts sound and picture together. J-cuts and L-cuts, where audio leads or trails the picture, need separate controls.
- A word boundary is safe, not always good. Splicing mid-phrase joins two pitch contours, which sounds edited even when no audio is damaged. The best cuts land in pauses.
- The picture still jumps. Removing a sentence from a static shot creates a jump cut that needs b-roll or a punch-in.
Common Misconceptions
- “Correcting the transcript changes what was said.” A spelling fix corrects text and captions, not the recorded voice. Generating new speech is a separate, generative feature.
- “Deleting a word deletes exactly that word.” It deletes a media range resolved from estimated timing, often including the gap beside it and rounded to whole frames.
- “The timestamps are exact.” They are model estimates with tens of milliseconds of error. The rendered splice, not the decimal, is the test.
- “It replaces the timeline.” For choosing dialogue, often. For timing, layers, audio and visuals, every complete editor still needs other controls.
How Valmera Uses It
Valmera is an agentic video editor that takes the index from text-based editing without the delete-to-cut interface. Every upload is transcribed once with word-level timestamps and speaker labels, alongside silence detection, shot detection and frames. You describe the edit (“cut the sentence about the old price,” “remove every ‘to be honest’,” “move the launch story to the start”) and the agent finds the words, cuts on word boundaries, checks the transcript the edit keeps, and renders a preview. The transcript is the map; the chat is the instruction.
The transcript panel shows every word, dims the ones the edit has cut, jumps the player when clicked, and lets you correct misheard lines, which updates the captions. Deleting text there does not cut footage. For hand edits, the timeline splits, trims and deletes segments. Cuts are stored in source seconds in an edit decision list, so every one can be restored. Captions are burned in, with no SRT or VTT export, and English is the best-tested transcription language. See text-based editing in Valmera.
Related Terms
- Word-level timestamps — The per-word timing the whole workflow is built on.
- Transcription and ASR — Where the text comes from, and the errors it carries.
- Edit decision list — What a transcript deletion actually writes: source ranges, not pixels.
- Silence detection — Words say what survives; pauses say where a cut can hide.
- Jump cut — What a transcript cut looks like on a single-camera shot.
- J-cut and L-cut — Split edits a single text selection cannot express.
Frequently Asked Questions
Sources
Made with Valmera: real exports
Edit by the Words, Without the Deleting
Valmera indexes your footage word by word, then edits it from a sentence. Creating an account and uploading are free; editing requires a subscription.
Create an account →