Accuracy depends on the audio, so measure it instead of trusting a vendor figure. Transcribe a three-minute sample, count the wrong words and divide by the total (the word error rate), and count names, numbers and negations separately. Clean narration needs a light edit, interviews a full one. Whatever the score, a person watches every file before delivery.
Vendor pages do quote figures. VEED's subtitle page claims accuracy of up to 99.9% and Kapwing's claims 99%, and neither explains what audio or method produced the number. VEED's page also suggests editing brand and industry terms to reach 100%. A production company needs its own numbers: a score per audio type, an error log, a glossary, timing limits, an export check and a watch-through.
Score a three-minute sample before you commit to a project
Word error rate, or WER, is the standard way to score a transcript. Add up the substitutions (a wrong word), deletions (a missed word) and insertions (an extra word), then divide by the number of words in the correct version. Three minutes of conversational speech is roughly 450 words, enough to show a pattern without costing an afternoon.
You need a reference, meaning the correct text. Make it by listening at normal speed while correcting the AI draft in a copy of the file, and ask a second person to check every name and number. Then run both versions through a document compare: Word has Compare on the Review tab, and Google Docs has Compare documents under Tools. A substitution shows up as a deleted word next to an inserted one, so count that pair once, not twice.
These are made-up results for three imaginary projects, shown to demonstrate the method, not to describe any tool:
| Sample | Audio | Reference words | Substitutions, deletions, insertions | WER | Critical errors | Route |
|---|---|---|---|---|---|---|
| A | Studio narration, one speaker, 2 min | 310 | 2, 1, 0 (3 in all) | 1.0% | 0 | Light check |
| B | Two-person interview, quiet room, 3 min | 480 | 12, 4, 3 (19 in all) | 4.0% | 5 | Full edit pass |
| C | Outdoor vox pops with wind, 2 min | 260 | 21, 8, 2 (31 in all) | 11.9% | 9 | Time the AI draft against human transcription |
Sample B works out as 19 errors divided by 480 words, which is 3.96%, so about 4.0% or 96% word accuracy. A critical error is one that changes the meaning or gets a name, number or brand wrong. Five of the nineteen were critical: two names, two figures and one dropped "not".
Three samples per audio type is a fair start: the cleanest stretch, the worst stretch and a typical one. The worst matters most, because that is where the client's hardest minute lives. Compare your samples with each other and with your own pass marks, not with a brochure figure, since a vendor's number and yours measure different audio.
My working thresholds, which are not an industry standard: under 2% on the raw draft, a light check; 2 to 8%, a full edit pass; above 8%, time one minute of fixing the draft against one minute of human transcription before choosing. Critical errors get fixed whatever the score. Set your own marks after ten files. WER also ignores two things that fail deliveries. It counts a wrong price the same as a wrong "the", and it says nothing about timing, which gets its own checks further down.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowNine error types to log, and the ones that must be zero
A single percentage hides what went wrong. Keep a log with one row per error: timecode, type, what the tool wrote, the correct text and a critical flag. After ten files the log shows which types your audio produces most, and for many productions it is names and numbers.
| Error type | Illustrative example | How to catch it | Critical? |
|---|---|---|---|
| Names and brands | A product name split into two ordinary words | Search every glossary term and its likely misspellings | Yes |
| Numbers, prices, dates | "Fifteen" written for "fifty" | Check each figure against the script or slide; search for digits | Yes |
| Negation flips | "We can deliver that by Friday" for "We can't deliver that by Friday" | Search "can", "not", "n't" and "no", and listen at each hit | Yes |
| Near-sounds and homophones | "Affect" for "effect" | A spell-check will not flag these; read for sense | Only if the meaning changes |
| Dropped words | A fast clause missing from a cue | Compare with the script; replay quick passages at normal speed | Depends on the clause |
| Text over silence or music | A phrase appearing where only a music bed plays | Scan cues that overlap music-only stretches of the timeline | Yes |
| Wrong or merged speaker | Two speakers' sentences in one cue | Check every speaker change; vital for captions with labels | Yes for captions |
| Punctuation and sentence breaks | A question rendered as a statement | Read the questions aloud | Sometimes |
| Offensive or off-brand misrecognition | An ordinary word rendered as a profanity | Search a short list of risky strings before delivery | Yes |
Negation flips are the quietest failure. Consider a testimonial about a delayed order, where the speaker says "we can't deliver that by Friday" and the draft reads "we can deliver that by Friday". Nothing flags it, the word-level score barely moves, and the producer notices only because the picture shows an empty loading bay. The defence is mechanical: search for every "can", "not", "no" and contraction ending in "n't" and listen once at each hit. It takes five minutes on a twelve-minute film.
The text-over-music row deserves its own pass. Speech models sometimes write words for silence or a music bed, and an invented line on screen is worse than a missing one. Jump to each stretch where the timeline carries only music and make sure no cue sits on top of it.
A names-and-jargon list built in ten minutes
Most critical errors are names and terms, and you know them in advance. The brief, the script, the slides and the client's product pages contain nearly all of them. Ask an assistant to pull them out so you can check the draft against a list. Use a business plan or turn off training first, because a client's unreleased material is going in.
From the material below (client brief, script, slide text, product pages),
list every item a speech-to-text tool is likely to get wrong: people's names,
company and product names, acronyms, technical terms, units and number formats.
Output a table with four columns:
Correct form | Likely misheard as | Where it came from | Notes (capitals, hyphens, spoken form)
Do not invent terms that are not in the material.
MATERIAL:
[paste]For a 12-minute interview with a facilities-software client, an illustrative result might read:
| Correct form | Likely misheard as | Source | Note |
|---|---|---|---|
| SaaS | "sass" | Script | Capital S at both ends, lowercase aa |
| HVAC | "H back" | Interview topics | Spoken as letters |
| SLA | "Sla" or "slaw" | Slide 6 | All capitals |
| [Client product name] | Two common words | Brief | One word, capital P in the middle |
| 3D | "three D" | Script | Numeral and capital D |
| [Speaker surname] | Several spellings | Speaker list from the client | Confirm with the speaker's own materials |
Review the list yourself before using it: assistants occasionally add a term the material never contained, which is why the prompt forbids it. Then use it twice. After generation, run a whole-word, case-sensitive find-and-replace for each "misheard as" variant you saw. Before generation, load the terms into the tool if it has a custom vocabulary field. VEED's and Kapwing's pages both describe correcting terms by hand in the transcript and neither describes a glossary feature, so check what yours offers and plan to fix by hand when it offers none.
Keep a running "misheard as" list per client. The second film for the same client starts with the corrections the first one taught you. Also settle house style on the same sheet: whether numbers under eleven are spelled out (Netflix's English guide spells out zero to ten and uses numerals above that), how job titles are capitalised, and whether the product name is written as the client writes it.
Line length, reading speed and cue timing, with the arithmetic
A subtitle file can be 99% right word for word and still fail delivery, because transcription and timing are separate jobs. The tool turns speech into words, then cuts the words into cues and times them. Word error rate measures the first job only. Netflix publishes a detailed timed-text style guide for many languages, which makes its English one a convenient default when a client gives you no spec. If the client or platform sends its own spec, use that.
| Limit | Netflix English guide | How to check it |
|---|---|---|
| Characters per line | 42 | Subtitle editor warnings, or a count on the longest lines |
| Lines per cue | 2 maximum, one line preferred unless the text will not fit | Scan for three-line cues in the cue list |
| Reading speed | Up to 20 characters per second for adult programmes, 17 for children's | Characters divided by display seconds |
| Minimum duration | Five-sixths of a second (20 frames at 24 fps) | Sort cues by duration, shortest first |
| Maximum duration | 7 seconds | Sort cues by duration, longest first |
| Gap between cues | At least 2 frames; gaps of 3 to 11 frames are closed to 2 at 24 fps | Run the editor's gap check |
Work one cue through. The text "We cut onboarding from three weeks / to four days across every site." has two lines of 34 and 31 characters, so the line limit is fine. Together that is 65 characters (66 if you count the line break as a space; subtitle editors usually count spaces and punctuation, so confirm in the spec you deliver to). If the cue stays on screen for 3.0 seconds, the reading speed is about 22 characters per second, over the 20 limit. Extending the cue to 3.4 seconds brings it to about 19. If the next cue starts straight after, you cannot extend, so condense the wording instead, and ask the editor or client before cutting a phrase that carries information, such as "across every site".
Line breaks matter too. Netflix's guide says to break after punctuation or before a conjunction or preposition and not to split an article from its noun. So "Our team reviews every file before the / client sees it." should read "Our team reviews every file / before the client sees it." A tool that cuts at a character count alone will produce the first version.
On the 12-minute interview film, the first pass had 168 cues. A scan found 23 above 20 characters per second (13.7%), 9 with a line over 42 characters (5.4%) and 6 shorter than five-sixths of a second (3.6%), 38 flagged cues in all. At about half a minute each, fixing them took roughly 19 minutes. These are illustrative figures. Your own first pass will show your tool's pattern, so count it, and check the segmentation settings (maximum characters per line, lines per cue) before you generate, where the tool offers them.
For the scan, Subtitle Edit is a free, open-source editor for Windows, macOS and Linux. Its Tools menu has Fix common errors, and its help pages show timing rules that can be set to Netflix's limits. Enter your spec's line length, reading speed and duration values, run the check and read the report, rather than assuming it catches everything. Always view the result in a player as well. A file that passes every automated rule can still read badly.
Subtitles or captions: speaker labels and sound cues
Ask the client which they are buying. W3C's guidance on captions says they identify who is speaking and include meaningful sound effects, while dialogue-only subtitles do not. A file meant for viewers who cannot hear the audio needs speaker labels and sound cues. A subtitle file for viewers who can hear the audio and want the dialogue in text does not.
For captions, Netflix's English style puts sound effects and speaker identifiers in square brackets, lowercase except for proper nouns, and does not italicise them. A before-and-after for a short exchange shows the difference:
Subtitles (dialogue only)
Cue 1: Are you coming?
Cue 2: Give me a minute.
Captions (dialogue, speakers, sounds)
Cue 1: -[host] Are you coming?
-[guest] Give me a minute.
Cue 2: [phone buzzing]Automated speaker separation is one of the weaker parts of the draft, so check every speaker change by ear. Use roles such as host and guest unless the client supplies names in writing.
Export formats and the file-level checks that catch drift
Choose the tool by where the editable text lands and what it exports, not by a headline accuracy number. Both VEED and Kapwing export subtitle files as SRT, VTT or TXT, or burn the subtitles into the video. On cost, Kapwing's page lists a Free plan with 10 minutes of auto-subtitling and watermarked exports, and Pro at $16 with 1,000 minutes a month (check whether that price assumes annual billing). VEED's pricing page lists Creator at $10 per user a month billed yearly, with 120 hours of subtitles a year, and Pro at $19 per user a month billed yearly, with unlimited auto subtitles under a fair-use limit of 50 hours a month. Its subtitle tool page puts Creator at $20 a month, so ask which billing period a quote assumes. Prices change, so verify before buying.
The two common sidecar formats look alike but differ in small ways that break files:
SRT
1
00:00:01,000 --> 00:00:03,500
We cut onboarding from three weeks
to four days across every site.
WebVTT
WEBVTT
00:00:01.000 --> 00:00:03.500
We cut onboarding from three weeks
to four days across every site.SRT is plain text with a cue number, a comma before the milliseconds and little support for styling. WebVTT must start with the word WEBVTT and uses a full stop before the milliseconds, and web players generally expect it. Streaming platforms often want TTML or IMSC, which are XML formats. Netflix's guide gives technical rules for them, such as percentage values instead of pixels, and points to its current delivery spec, so never assume a TTML file exported from a general tool will pass.
A burned-in (open) subtitle is part of the picture and cannot be corrected without exporting the video again. If you deliver both a burned-in version and a sidecar file, check them separately.
Four file-level checks catch most of the problems that survive a good edit:
- Encoding. Open the file in a plain-text editor. Accented letters and curly quotes should look right. If "é" appears as "é", the file was saved in the wrong encoding, so save it again as UTF-8.
- Offset and frame rate. A sidecar file is right only for the picture it was timed against. A different start timecode, a colour-bar and tone lead-in, edits made after timing, or a conversion between frame rates can make cues land early or late. Check the first cue, one in the middle and the last against the final master.
- Naming. Put the language in the file name, for example filmtitle_en.srt, and use the same pattern on every file in the delivery.
- Load test. Open the exported file, not the project, in the player the client will use, with the subtitles switched on, and watch the first minute.
The QC pass on a 12-minute film, minute by minute
Here is the whole routine for the interview film, timed. The figures are illustrative, and the "by hand" comparison assumes your editor now spends about four hours subtitling a film this length. Replace both with your own timings from a pilot.
| Step | Minutes |
|---|---|
| Upload, generate, download the draft | 4 |
| Glossary find-and-replace | 6 |
| Edit pass, listening at normal speed | 30 |
| Fix the 38 flagged timing and line-length cues | 19 |
| Automated check in Subtitle Edit | 5 |
| Full watch-through of the film | 12 |
| Fixes from the watch-through | 8 |
| Export, encoding and sync spot-checks | 6 |
| Total | 90 |
Ninety minutes against four hours is a saving of 150 minutes on this film. The watch-through is the step to protect when time is short, because it is the only one where a person sees picture, sound and text together. Watch at normal speed with the sound on. Look for cues that cover on-screen text or lower thirds, for flashes of text across shot changes, for the first and last cue, and for credits.
The pass has a failure rule. Any critical error sends the file back. So does a watch-through that finds more than three minor errors in ten minutes, which is my own rule of thumb and tells you the edit pass was too light. Record the result on a short sign-off sheet:
FILM: [title] v3, running time 12:04
LANGUAGE: en | FORMAT: SRT sidecar plus burned-in social cut
SPEC: client house style with Netflix-style limits (42 / 2 lines / 20 cps)
GLOSSARY APPLIED: yes, 14 terms
SAMPLE WER BEFORE EDIT: 4.0% | CRITICAL ERRORS FOUND AND FIXED: 5
AUTOMATED CHECK: Subtitle Edit, Fix common errors, 0 open
SYNC SPOT-CHECKS: 00:12 ok | 06:02 ok | 11:48 ok
WATCH-THROUGH: [initials], [date], 3 minor fixes made
NAMES AND NUMBERS CHECKED AGAINST SCRIPT: yes
ENCODING: UTF-8 confirmed
SIGNED OFF: [initials]If a film will also be translated, finish and sign off the source-language file first. Errors that survive into the master are copied into every translation, and the checks for that stage are in the tutorial on AI subtitle translation quality checks. The same logic applies to dubbing: see AI video translation for corporate video.
What to tell the client about accuracy
Say what was done in the delivery note, in two sentences. For example: "Subtitles were drafted with AI, then corrected by our team against the audio and checked in a full watch-through. Names, figures and technical terms were checked against your script and glossary." Add the file formats and the spec the limits follow.
You will often be asked how accurate it is. Give your own sample score for that audio, not a vendor figure, and name the checks that sit behind it. A client who sees a per-project number and a signed sheet trusts the next delivery more than one who is shown a headline percentage.
Two related habits pay for themselves. Run the sample test again whenever you switch tools or the audio type changes, and see the tutorial on testing a transcription tool on your own calls for the same idea applied to phone audio. And keep the corrected transcript, because the same text feeds rough cuts, as in building a rough cut from a transcript.
Subtitle QA questions production teams ask
What word error rate is acceptable for client subtitles?
There is no universal figure. As working thresholds, I would treat under 2% on the raw draft as a light check, 2 to 8% as a full edit pass, and above 8% as a prompt to time human transcription against fixing the draft. Whatever the raw score, the delivered file should have no critical errors in names, numbers or negations.
Do I have to say the subtitles were AI-generated?
Say so in the delivery note, in plain words: AI-drafted, checked by a person against the audio. Some client specs and platforms set their own rules on automated captions, so read the spec you are delivering to and ask the client if it is silent. Disclosure costs one sentence and avoids an argument later.
Can free subtitle tools be used for client work?
Only for a test. Kapwing's free plan limits auto-subtitling to 10 minutes and watermarks exports, and VEED's page says longer videos and subtitle file downloads need a paid plan. Also read each tool's data terms before a client's unreleased footage goes in, since some tools use uploads for training unless you switch it off.
Should subtitles be verbatim or cleaned up?
Agree it before generating. Most client films use clean verbatim, dropping false starts and filler but keeping every meaningful word. Legal, research or training-evidence videos may need true verbatim. Whichever you choose, write it in the style sheet so the AI draft, the editor and the QC pass all work to the same rule.
Further reads
- AI vs Human Transcription: Which Is Worth Paying For? — When a human transcriber beats fixing an AI draft.
- AI Tools for TV Production: Transcripts, Logging and Clearance Notes — Transcripts and logging on longer productions.
- Subtitling Rates per Minute: What to Charge When AI Drafts First — Pricing the human pass when AI drafts the subtitles.
- AI Voice Over for Videos: Can a Production Company Use It for Clients? — Consent questions when AI voices are added to client films.
- Best AI Filmmaking Tools by Job for a Small Production Business — Tools for paperwork, rough cuts, audio, captions, voice and generated shots, with plan costs and the check to run before a client sees the file.
- Video Production Proposal: Scope, Budget and Timeline Drafted With AI — A worked proposal for a three-minute corporate film: prompts for each section, the budget sums, revision-round wording and payment milestones.
- AI Video Content Agency: Services a Small Studio Can Add and Price — A six-service menu for a small video studio, a worked pricing floor, and the rights, consent and disclosure checks before delivery.
- AI Video Production Software Compared for a Small Studio — Tools grouped by script, edit, captions, voice, shorts and generated shots, with a three-editor monthly cost and licensing limits.
- Video Production Pricing Sheet: Build a Rate Card With AI — A rate card for video work built from your own costs: day rates, edit hours, revision rules and licence fees, with a worked brand-film quote.
- AI Shot List Generator: Build a Shoot-Day Plan From the Script — A worked 90-second corporate film from script to shot list, schedule and call sheet, with the director's checks that catch what AI misses.
- AI Storyboard Generator From a Script: A Video Producer's Workflow — A beat-by-beat workflow for turning a client script into a storyboard with AI, plus a test for the frames that still need a human artist.
- AI Video Script Generator for Production Companies: Brief to Draft — Use a two-pass brief-to-script process that protects client facts, runtime and approval history.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Netflix partner help timed-text style guides (general requirements, English, timing guidelines), W3C Understanding WCAG captions page, VEED and Kapwing subtitle and pricing pages, Subtitle Edit documentation (checked November 2026).