AI Transcription Workflow for Freelance Video Subtitling

Machine transcription has quietly become the default first step in video production. What used to take hours of typing now happens while you make coffee, which means the interesting question has changed: not whether an engine can turn audio into words, but whether the words survive contact with a phone screen, a search result, a translated track, and an accessibility audit. This guide walks through a practical, tool-agnostic workflow for transcription and subtitling, including where machine output is dependable, where it quietly fails, and how to build quality checks that stop rework before it starts.

What AI Transcription Does Well — and Where It Still Fails

Modern speech-to-text engines are genuinely good at the easy 70% of the work and consistently shaky on the 30% that clients notice. Knowing which is which lets you plan review time honestly instead of promising a clean deliverable on a hopeful assumption.

Reliable wins

A 40-minute, two-person interview on decent audio usually returns a usable draft in one to three minutes.
Punctuation restoration, capitalization, and paragraph splitting are strong enough to produce a readable transcript with light editing.
Speaker separation (diarization) works well when two voices are distinct and not talking over each other.
Word-level timestamps make it possible to build karaoke-style captions, jump-cut edits, and chapter markers without manual scrubbing.

Predictable weak spots

Overlapping speech: when two people laugh or agree at the same time, engines often merge them into one speaker or drop one entirely.
Domain vocabulary: product names, internal acronyms, medical terms, legal phrasing, and invented brand spellings.
Homophones and near-homophones: their/there, cache/cash, phase/faze, and names that sound like common words.
Numbers, dates, units, and ranges, which are exactly the details clients check first.
Music beds, ad reads, phone-line compression, and room echo, which push error rates up sharply.
Code-switching: a sentence that starts in one language and ends in another often produces gibberish in both.

The operational takeaway is simple: treat machine output as a draft with unknown error locations, not as a finished file. Your review pass is not a formality, it is the product.

Mapping the End-to-End Subtitle Workflow

A repeatable four-stage pipeline keeps transcription work predictable, whether you handle a single interview or a weekly series.

Stage 1 — Intake and preparation

Ask for the original recording rather than a re-encoded export from a phone. A mono 48 kHz file with stable loudness and no clipping gives you the best possible starting point. Before running anything, write down the glossary (names, brands, jargon), the speaker names, the number of speakers, the target languages, the delivery formats, and the project frame rate. Five minutes of setup here saves an hour later, because most timing drift and spelling churn trace back to missing context.

Stage 2 — The machine pass

Run the engine with the custom vocabulary list loaded, the language hint set correctly, diarization enabled, and word-level timestamps on. Export two separate artifacts: a raw text transcript with paragraphs, and a timestamped caption file. Keeping them separate matters, because editing a transcript for readability and editing captions for timing are different jobs with different rules.

Stage 3 — Human review and timing

Do a read pass first (silently, at full speed) to catch obvious nonsense, then a listen pass at 1.25x speed with the audio and captions side by side. Fix names and numbers against the glossary, split long cues, and move in/out points to natural pauses rather than mid-word. If a caption appears noticeably before or after the spoken words, viewers feel it as sloppiness even if they cannot name the problem.

Stage 4 — Export, delivery, and archiving

Deliver in the formats the client actually needs: SRT or WebVTT for web players, TTML or IMSC for broadcast and streaming pipelines, a clean transcript for publishing, and ideally a plain text version for reuse. Use a consistent filename convention that includes the video ID, language, and format. Archive the source audio, the glossary, and a note of which engine version produced the draft, so a future revision is a diff rather than a redo.

Building a Tool Stack Without Brand Lock-In

No single engine wins every job. A stack that mixes local processing and cloud services gives you options for confidentiality, speed, and language coverage without depending on one vendor's roadmap.

Local versus cloud processing

Legal, medical, and pre-release marketing material often cannot leave a controlled environment, which makes offline engines the practical choice despite slower throughput. Cloud services win on speed, batch processing, language breadth, and translation add-ons. A sensible compromise: process sensitive material locally and push public-facing marketing content to a cloud batch job.

An evaluation checklist

Run candidates against a ten-minute benchmark clip that includes your hardest conditions: two speakers, one accent, one jargon-heavy passage, and background music. Score each engine on:

Word error rate on that clip, measured against a hand-corrected reference.
Diarization quality when speakers interrupt each other.
Word-level timestamp granularity and stability.
Custom vocabulary and hotword support.
Export formats and whether timestamps survive export.
Batch or API access for long series work.
Language coverage for the regions you serve.
Privacy terms, retention policy, and whether processing can happen offline.
Processing cost per minute at your typical volume.

When to combine two engines

For high-value or noisy projects, run two engines and compare the outputs. Agreements are almost always correct; disagreements point directly at the hard parts of the recording — which is where a human reviewer should spend time anyway. This consensus approach is slower but it converts a fuzzy review process into a targeted one, and it is especially effective for names, numbers, and technical terminology.

Quality Control Checkpoints That Prevent Rework

Quality control for captions is not subjective. Set numeric thresholds, then check against them.

Accuracy targets

For clean studio audio with a loaded glossary, an edited deliverable should have effectively zero errors in names, numbers, and brand terms. General word accuracy in the high nineties is achievable with review. A practical sampling rule: review the first five minutes completely (that is where unfamiliar jargon clusters), then spot-check ten to twenty percent of the rest at random, plus any timestamp flagged by the comparison pass.

Reading speed and line breaks

Captions fail when they are technically accurate but impossible to read. Working defaults:

15 to 20 characters per second reading speed.
Maximum two lines per cue.
Around 42 characters per line.
Minimum cue duration of one second, maximum of six to seven seconds.
A short gap between cues, roughly two frames, so the eye can reset.
Break lines at clause boundaries, and never split a name, number, or article-plus-noun pair across two cues.

Speaker labels and sound cues

Label speakers on first appearance and whenever the speaker changes after a gap. Include meaningful non-speech audio in square brackets — [music], [laughter], [door closes] — but do not narrate everything; if it does not change the meaning, it does not belong in the caption. Use italics or a distinct label for voiceover and off-camera narration, and keep that convention identical across every episode of a series.

Captions as Search Assets

Caption files and transcript pages are indexed, quoted, and summarized by search engines and AI assistants. That makes transcription accuracy a discovery issue, not only an accessibility one.

Start with correct text

A misheard product name is invisible on video and fatal in search. If your captions say the wrong word for your central topic, no amount of metadata will fix the mismatch. This is the strongest practical argument for reviewing the passages that carry your key phrases before publishing anything.

Transcript pages and structure

Publish transcripts as accessible HTML with headings, question-and-answer structure for interviews, and timestamps where they help navigation. Pair the page with structured data that points to the video and its caption track so search systems can associate the text with the media object. Keep the transcript human-readable first: if it reads like keyword soup, viewers leave and the page underperforms anyway.

Sync chapters with caption boundaries

Align chapter timestamps to sentence boundaries rather than arbitrary seconds, and put the primary topic phrase in the chapter title. Chapters that land mid-sentence feel broken, and broken navigation reduces watch time on exactly the content you want people to finish.

Accessibility and Multi-Language Delivery in Plain Terms

Accessibility compliance is often described in legal language that hides a simple checklist.

The essentials

Provide captions for all pre-recorded spoken content, and a transcript for people who cannot use captions at all. Identify speakers, because who said something is often as important as what was said. Maintain strong contrast between caption text and background, use a readable font size, and never rely on color alone to convey meaning. Avoid burning captions into the master export when a sidecar or embedded track will do, since burned-in text cannot be edited, translated, or restyled later.

Translating captions without breaking timing

Subtitle translation is not document translation. Character counts expand or contract dramatically by language, and reading speed differs. Translating the text and swapping it into the existing cues produces caption tracks that flash by or sit on screen far too long. The reliable sequence is: finalize the source captions, translate with the timing constraints visible, then re-time the target track and re-run the reading speed check. Budget for expansion in one direction and contraction in the other, and consider a native reviewer for anything customer-facing.

Managing multiple tracks

Use a consistent naming convention that includes the video identifier, language code, and format, and decide early whether tracks are embedded, sidecar, or burned in for vertical social edits. Keep a single source of truth for the glossary in every language so spellings do not drift between episodes.

Comparing the Practical Approaches

Approach	Best for	Review effort	Main risk
Manual transcription only	Short, highly sensitive, or heavily accented clips	High	Slow throughput and fatigue errors
Machine draft plus full human review	Client deliverables, published transcripts, series work	Moderate	Time underestimated on noisy audio
Machine draft plus sampled review	Internal reference, rough cuts, archive search	Low	Rare but real errors in unchecked sections
Machine output published as-is	Private notes, personal archives	None	Names, numbers, and compliance failures

The middle two rows are where most professional work lives. The deciding factor is not the length of the video but the cost of an error: if a wrong number changes a decision, budget for a full review.

Scoping and Delivering Subtitle Work for Clients

Good scoping turns transcription from an unpredictable favor into a deliverable you can estimate.

Deliverable tiers

Define tiers explicitly so expectations are set before work starts. A raw machine transcript is tier one. An edited, readable transcript with speaker labels is tier two. Timed captions with reading-speed compliance and non-speech cues is tier three. Localized caption tracks with a quality report and glossary is tier four. Most disputes come from a client assuming tier three while the agreement described tier one.

Turnaround and revision policy

Agree on a glossary and speaker list before the first pass, freeze the audio when review begins, and state plainly what counts as a revision versus a new request. For planning, assume a clean 20-minute interview takes roughly 25 to 40 minutes of review and timing work; heavily accented or crosstalk-heavy material can double that.

Handoff documentation

Ship a short note with every delivery: engine and version used, glossary applied, caption style settings, formats included, and any passages that needed a judgment call. This single page reduces follow-up questions dramatically and makes the next episode easier to plan.

Common Mistakes and How to Avoid Them

Publishing unreviewed machine output. Fix: always run at least a targeted review pass on names, numbers, and key phrases.
Ignoring reading speed. Fix: run a compliance check for characters per second and cue duration before export.
Translating before timing is final. Fix: lock source captions first, then translate and re-time.
No glossary. Fix: build a shared term list per client and load it into the engine before the first pass.
Treating captions as an export-time afterthought. Fix: decide formats, styling, and languages during intake.
Losing sync on non-standard frame rates. Fix: confirm the project frame rate and verify drift at the end of long files, not only at the start.
Deleting source audio after delivery. Fix: archive audio, glossary, and engine notes together for fast revisions.
Burning in captions on the only master. Fix: keep a clean master plus a separate caption track.

FAQ

How accurate is AI transcription today?

On clean single-speaker audio it is close to publishable. On interviews, accents, and technical vocabulary, expect a usable draft containing a small number of concentrated errors — which is why targeted review, not blind trust, is the standard practice.

Can automatic captions meet accessibility requirements?

They can be a starting point, but unchecked output routinely mislabels speakers and garbles names. Review for speaker identification, reading speed, and non-speech cues before treating any track as compliant.

Should I deliver SRT, WebVTT, or both?

Deliver both unless the platform is fixed. SRT is widely supported but limited, while WebVTT handles styling and positioning better. For broadcast pipelines, add TTML or IMSC.

How do I handle overlapping dialogue?

Split the overlap across consecutive cues with speaker labels rather than stacking text, and keep each line short. If the overlap is unintelligible, describe it rather than inventing dialogue.

How long does reviewing an hour of audio take?

For clean content, plan on one to two hours of careful review and timing; add time for accents, crosstalk, multiple languages, or dense technical vocabulary. Machine speed does not remove this step, it only shifts the effort.

Do I need word-level timestamps?

They are essential for karaoke captions, precise editing, and chapter placement, and useful for locating errors quickly. For plain transcripts they add little, so choose tools that let you switch them on per project.

Short vertical clips benefit from larger text, fewer words per cue, and slightly slower reading speed. Keep the master caption track separate so a social reformat never damages the original deliverable.

The teams that get the most from machine transcription are not the ones chasing the newest engine. They are the ones with a written intake checklist, a glossary per client, a review pass with numeric thresholds, and an archive that makes revision cheap. Pick your engine against your own benchmark clip, keep a second option for hard jobs, and treat the final ten percent of human judgment as the part that actually differentiates the work.