Dialogue trimming: cut before voiceover
This page polishes the dialogue column of Document Two: the short drama script. In 2.2 the short drama script template you wrote a first draft of dialogue for every beat—2.2 allows this step to be written loosely; get the meaning right first. Short-video dialogue writing accepts only one hard metric before voiceover: the entire film's dialogue can be read aloud within 10 seconds—this is the line drawn for Document Two by the 2.2 dialogue line. If it exceeds that, cut—this page teaches you to cut with justification: do the math first, then run every line through the deletion line, and finally let your mouth do the acceptance check.
What you'll be able to do after this page
- Convert the word count of the script's dialogue column into voiceover duration and subtitle line counts, and know the maximum amount of dialogue a 60-second piece can hold
- Use the three-question deletion line to decide line by line: cut what the audience can understand, what the picture can show, and what the character wouldn't say
- Rewrite "narrator-style" explanatory lines into visual actions, letting the camera speak for the characters
- Read aloud with a stopwatch, squeeze the entire film's dialogue into the 10-second budget, and fill the script back in to finalize Document Two
- When you can't cut any further, have AI act as a dialogue doctor and give line-by-line suggestions following the deletion line (a prompt designed for this course)
Your current predicament
- "The script felt satisfying when finished, but reading the dialogue straight through takes nearly a minute—a 60-second piece simply can't hold it"
- "Only after finishing the voiceover did I find the lines ran over; I had to re-record everything, and redo the subtitles too"
- "All the lines are narration: 'These steps are really slippery,' 'The cake is about to collapse'—things the audience can see, and I said them again"
When to use this technique
When you need to:
- Cut the loosely written dialogue in the 2.2 script down to "every line is necessary," so 60 seconds can genuinely hold it
- Lock down line length before voiceover—the same subtraction applies to both character dialogue and narration voiceover scripts
- Catch up on the fast-track route: you can do this lesson now, or save it for right before 5.1 voiceover (the fast-track route in the 2.0 guide)
And you don't want to:
- Discover only after recording that the dialogue crowded out the visuals, forcing a full re-sync of audio and picture
- Force lines in by speeding up the AI voiceover's speech rate, leaving the audience unable to hear clearly and the subtitles unable to keep up
🎒 Preparation before you start
- Completed 2.2 Short drama script template: one A4 page to lay out your 60 seconds, with the first draft of dialogue already in the "dialogue" column of the script table
- That one-page A4 script is at hand—do the subtraction directly in the "dialogue" column; no new file needed
- A phone or stopwatch that can time: step 4 requires reading aloud once
- No tools or GPU needed: this page is still paperwork; the AI voiceover's speech-rate parameter is only previewed in the "pitfall reminders," with hands-on work in 5.1 voiceover and licensing
Core idea
The unit of measure for dialogue isn't the "sentence," it's the "second". Every word on paper takes up two things: voiceover duration and subtitle line count. So the first step of subtraction isn't cutting—it's conversion: first know how many seconds each line is worth, then decide who stays.
Conversion relationships (experience values from this course, not official specs of any tool):
| Conversion | Experience value | A line of 20 characters |
|---|---|---|
| Word count → voiceover duration | Chinese voiceover about 3–4 characters per second | about 5–7 seconds |
| Voiceover duration → subtitle line count | about 10–15 characters per subtitle line in portrait mode | about 2 lines |
Lines with emotion or pauses (whispering, ellipses) eat seconds; estimate conservatively at 2–3 characters per second.
What's the dialogue budget for a 60-second piece? The 2.2 dialogue line has already drawn it: the entire film's dialogue can be read aloud within 10 seconds. Converted at 3–4 characters per second, that's about 30–40 characters. Why so few? In 60 seconds, visual actions, silence, and ambient sound are all telling the story—dialogue is the resource of last resort. Too many lines and the voiceover crowds the picture while subtitles smear across the screen.
Overlong dialogue discovered at different stages carries completely different costs (the course's accounting logic; the full ledger is in 1.4 cost and trial-and-error log):
| When the overage is discovered | Cost |
|---|---|
| On paper (this page) | Cut a few characters, zero cost |
| After voiceover | Re-record the whole segment, regenerate subtitles |
| After editing | Full re-sync of audio and picture; post-production essentially redone |
That's where this page's title comes from: subtraction must happen before voiceover—deleting one character on paper is the cheapest.
Follow along
The following example continues with the teaching example topic from 2.2: under a pedestrian bridge in a rainstorm, delivery rider Old Chen shields a birthday cake for his daughter. 2.2 allows the dialogue first draft to be written a bit loose—if your first draft has a line of dialogue in every beat, this lesson is the main battlefield; if it's already very short, treat the four steps as a review. You can also follow along with your own script throughout.
Step 1: Do the time math for short-video dialogue writing
Why: Dialogue length isn't a feeling, it's seconds. Without converting, you don't know how much you're over.
Count the characters of every line in your script's "dialogue" column (a rough count including punctuation is fine), total them for the whole piece, and apply the conversion table above. A loose first draft of the same story usually looks like this (teaching example, the "a bit loose" version allowed by 2.2):
| Beat | Dialogue first draft | Character count |
|---|---|---|
| 1 | Old Chen (to himself): "The cake mustn't get wet—almost home now." | 15 |
| 2 | Old Chen: "Why are these steps so slippery? What a rotten day." | 15 |
| 3 | Old Chen: "Oh no, oh no, the cake collapsed—what do I do? My daughter will be so disappointed." | 21 |
| 4 | Daughter (offscreen): "Daddy, why did the cake collapse?" Old Chen: "I'm sorry, Daddy took a fall." Daughter: "It's still pretty collapsed, like a snowy mountain." | 32 |
Total: about 83 characters. Converting at 3–4 characters per second: reading it aloud takes 21–28 seconds—in a 60-second piece, roughly forty percent of the time is spent reading lines, squeezing out the visuals and the silences.
What you should see: A character-count column appears next to the script's dialogue column, with a total and converted seconds at the bottom. Whether you're over budget (10 seconds / 30–40 characters) is visible at a glance.
- Small check: the total dialogue character count for the whole piece is calculated, and the converted seconds are written in the script's margin
Step 2: Run the three-question strikethrough—ask three questions about every line of dialogue
Why: You can't cut dialogue by feel—feel will mistakenly delete emotional landing points and let filler slide. Run every line through the same three questions, so every cut or keep has a basis (this course's teaching template):
The dialogue three-question strikethrough
□ Information question: If you delete this line, can the story still be understood? If yes → delete
□ Visual question: Can the camera act out what this line says? If yes → delete, and move the information into the visuals column
□ Authenticity question: Would this character, at this moment, actually say this line out loud? If not → deleteTake the first draft from Step 1 through it line by line:
| Line | Three-question result | Verdict |
|---|---|---|
| "The cake mustn't get wet—almost home now." | Visual question: him clutching the box to his chest while sprinting says it all; Authenticity question: someone sprinting has no breath to narrate information to himself | Delete |
| "Why are these steps so slippery? What a rotten day." | Visual question: the stumble from his slipping soles speaks for itself | Delete |
| "Oh no, oh no, the cake collapsed—what do I do? My daughter will be so disappointed." | Information question: the audience can see the half-collapsed cream; "my daughter will be disappointed" is psychology the camera can't shoot | Compress |
| "Daddy, why did the cake collapse?" | Information question: it asks about the collapse the audience just watched; this question exists only to set up the recap in the next line | Delete |
| "I'm sorry, Daddy took a fall." | Information question: it recaps the fall the audience just saw in the previous beat | Delete |
| "It's still pretty collapsed, like a snowy mountain." | The emotional landing point of the whole piece; the third-line ending in 2.1 lands on it | Keep |
Only lines that pass all three questions earn their place; lines marked "compress" go to Step 3 for handling.
What you should see: Every line of dialogue now has a verdict next to it—delete, compress, or keep—and every reason traces back to the three questions, not to "I feel like it."
- Small check: for every line kept, you can say which question let it through
Step 3: Replace explanatory lines with visual action
Why: Of the three questions, the one "narrator lines" fail most easily is the one about restating what the audience can already see. But deleting doesn't mean losing the information: the information must land in the script's "visual action" column.
Side-by-side rewrite (teaching example):
| Explanatory line ❌ | Landed in visual action ✅ |
|---|---|
| "Why are these steps so slippery" | His soles slip on the wet steps and he stumbles (already in the visuals column) |
| "Oh no, the cake collapsed" | He stares at the half-collapsed cream, fingers hovering over the box lid, not daring to touch it |
| "I'm sorry, Daddy took a fall" | He wipes the rain off his face, opens his mouth at the phone, and says nothing |
Compression is also a kind of rewrite: "Oh no, oh no, the cake collapsed—what do I do? My daughter will be so disappointed" is 21 characters, compressed to "Don't collapse, don't collapse…"—no restating the visuals, only the pleading, and not a bit of emotion lost.
What you should see: The dialogue column contains only lines that passed all three questions; the information from deleted lines can all be found in the visual action or sound columns.
- Small check: verify each deleted line—no information was lost; either the visuals already had it, or it's been added to the visuals/sound columns
Step 4: Read it aloud once with a stopwatch, then fill in the final script
Why: The conversion table is an estimate—pauses and emotion don't consume character counts, they consume seconds—your mouth is the final judge.
Read the trimmed dialogue through once at a normal speaking pace, with a stopwatch:
Old Chen (in a low voice): "Don't collapse, don't collapse…"
Daughter (offscreen): "It's still pretty collapsed, like a snowy mountain."12 characters (excluding punctuation), about 4 seconds to read aloud—within the 10-second budget, with room to spare for the low-voice pauses and the phone-connecting tone. This is the final version of the dialogue column in the 2.2 script: only two lines of dialogue in the whole piece—not laziness, but what it looks like when the subtraction is done.
The before-and-after accounting of the subtraction (all conversions are course experience values, not official specs of any tool):
| First draft | Final draft | |
|---|---|---|
| Dialogue character count | About 83 characters | 12 characters (excluding punctuation) |
| Voiceover duration | About 21–28 seconds | About 4 seconds |
| Subtitle line count | About 6–9 lines | 2 lines |
| left for visuals and silence | 35 seconds or so | 55 seconds or so |
Fill the finalized dialogue back into the script's "Dialogue" column, and while you're at it, mark each line with its estimated seconds—5.1 Voiceover and Licensing records against it, and 5.2 Subtitle Production aligns against it. File Two is finalized here.
You should see: the script's dialogue column replaced with the finalized version, each line carrying its estimated seconds; read aloud once, it takes ≤10 seconds, and the dialogue line of 2.2 officially passes.
- Quick check: timing yourself reading it aloud finishes in ≤10 seconds; the script's dialogue column has been backfilled with the final version and annotated with seconds
Checkpoint ✅
- The whole film's dialogue can be read aloud within 10 seconds—actually timed with a stopwatch, not estimated
- Every retained line has passed the three-question deletion test, and you can state the reason for keeping it
- The information from deleted lines all has a destination: the visual action column or the sound column, with no information lost
- The script's "Dialogue" column has been backfilled with the final version, each line annotated with estimated seconds—File Two is finalized
- Overlength was not solved by "speeding up the speech rate"—the speech rate parameter is left for fine-tuning in the voiceover stage
Pitfall Reminders
❌ Forcing overlong dialogue in with the speech rate parameter
Symptom: after cutting all you can, the lines still have 80-plus characters and take 20-plus seconds to read—you don't want to cut more, so you plan to set the AI voiceover's speech rate to maximum and cram it into the 10-second budget.
Cause: the speech rate parameter is a fine-tuning tool, not a compressor. Past a certain speed, viewers can't hear clearly and subtitles can't keep up; and discovering overlength only at the voiceover stage makes it the most expensive to fix (see the cost table in the core ideas).
Correct approach: first cut the dialogue down to the 10-second budget on paper (the four steps on this page); in the voiceover stage, only make small adjustments to the speech rate. edge-tts does provide a speech rate parameter: --rate (e.g., --rate=-50% slows down; negative values must be written joined as --rate=-50%, not as --rate -50%), to be used only when you get to 5.1 Voiceover and Licensing.
❌ Overdoing the subtraction and deleting the emotional landing point along with it
Symptom: you got hooked on cutting, and not a single line remains in the whole film—even the daughter's line "It's beautiful even collapsed, like a snowy mountain" got cut.
Cause: "every line must be necessary" was misread as "the fewer, the better." Information can be carried by visuals, but the emotional landing point often can only be carried by dialogue—that closing line is how the third sentence of the 2.1 Story Synopsis (ending and emotional landing point) lands in sound.
Correct approach: before cutting any line, ask one extra question beyond the three: "If I cut it, is the ending's emotion still there?" The emotional landing point line is the last line in the whole film you'd consider cutting, and usually shouldn't be cut at all.
❌ Only counting characters, never reading aloud
Symptom: the table calculates 35 characters ≈ 10 seconds, but actually reading it aloud takes 15 seconds—two pauses and one quiet, slow line, none of it counted.
Cause: the conversion table gives averages for normal speech speed; markers like "quiet" and "……" all eat seconds, and numbers on paper can't see them.
Correct approach: before finalizing, you must read it aloud once with a stopwatch (step 4). Estimate emotional lines conservatively at 2–3 characters per second, leaving margin for pauses.
Advanced Tips (Optional)
When you can't cut any further, let the AI act as a dialogue doctor—same usage as the script doctor in 2.2 Script: ask for line-by-line diagnosis, not a rewritten final draft. Which lines to cut and which to replace with visuals—the AI gives suggestions, but you decide what stays.
Expected result: you get a line-by-line verdict list—cuttable, replaceable with visuals, must keep—each with a reason.
## Role
You are a dialogue editor for 60-second short videos, skilled at cutting lines down to "every line is necessary."
## Task
Check my short drama dialogue line by line and give a verdict based on the three-question deletion test: cut / compress / replace with visuals / keep, with a one-sentence reason for each.
## Input Information
### Required Fields
- Full dialogue text: [paste the script's dialogue column, with beat numbers and speakers marked]
- Total film duration target: [e.g., 60 seconds]
### Optional Fields
- Lines that must be kept: [e.g., the emotional landing point line at the end]
## Output Specification
Output line by line: original line → verdict (cut / compress / replace with visuals / keep) → one-sentence reason; at the end, give the total character count after cutting and the estimated voiceover seconds.
## Constraints
- ✅ Only give verdicts and reasons; do not rewrite the entire dialogue—what stays or goes is my decision
- ✅ Suggestions to replace with visuals must fall within the existing scenes and characters
- ❌ Do not add new characters, add lines, or alter the main storyline💡 The boundaries of AI assistance
For lines the AI judges as "cuttable," run them through the three questions yourself once more before cutting—it doesn't know where your emotional landing point is. After finalizing, still read aloud once with a stopwatch: the AI counts characters, your mouth counts seconds.
Lesson Summary
✅ At this point, you have:
- Converted the length of short-video dialogue writing into voiceover duration and subtitle line counts, setting a dialogue budget of 10 seconds / 30–40 characters for a 60-second film
- Made line-by-line trade-offs with the three-question deletion test, cutting a loose 83-character first draft down to a 12-character final draft
- Landed the information from explanatory lines in the visual action column, letting the camera speak for the characters
- Verified by reading aloud once with a stopwatch and backfilled the script—File Two finalized, giving both voiceover and subtitles an execution basis
Dialogue subtraction quick reference:
| Item | Value / Standard | Notes |
|---|---|---|
| Dialogue budget | Whole film read aloud ≤10 seconds (about 30–40 characters) | Consistent with the dialogue line of 2.2; course rule-of-thumb value |
| Speech rate conversion | 3–4 characters per second; 2–3 for emotional lines | Rule-of-thumb value, not an official spec of any tool |
| Three-question deletion test | Information question / Visual question / Authenticity question | This course's teaching template |
| Subtitle line count | About 10–15 characters per line in portrait mode | Rule-of-thumb value; line count follows character count |
| Speech rate parameter | edge-tts --rate=-50% (negative values written joined) | For fine-tuning in the voiceover stage, not a compressor |
Next Lesson Preview
The dialogue is finalized and File Two is complete. The next lesson breaks it into shots: Storyboard: Breaking the Script into a Shot List.
You will learn:
- Breaking the finalized script into a numbered shot list—File Three, with each shot individually ready to be taken for generation
- The storyboard's fixed fields: shot number, shot size, visual, sound, duration
- How the 3–5-second-per-shot slicing is determined (a course design value), and how the shot count links to the 1.4 Cost Table