A source script and a finished narration file may contain the same words but not the same timing. Forced alignment solves a narrower problem than speech recognition, and knowing that distinction makes captions both faster and more reliable.
Choose your starting point
If you have a verified script and its corresponding recording, forced alignment estimates when the known words were spoken. ElevenLabs documents an alignment API for that task and explicitly says diarization is not supported in that route. If you have audio but no reliable script, run transcription first and correct the text before aligning. Aligning an inaccurate script can produce deceptively precise timestamps for words that were never said.
Build the SRT or VTT
Normalize the spoken script, align audio against it and inspect every segment at speaker switches, long pauses and fast technical lists. Convert word timing into caption blocks with human-readable line breaks, adequate on-screen duration and punctuation. SRT uses numeric cue numbers and comma-separated millisecond times; WebVTT uses a WEBVTT header and period-separated millisecond times. Validate with a parser before delivering to a player or video editor.
Word accuracy comes before timestamp accuracy
In a historical documentary, the difference between 1857 and 1957 is editorial, not just phonetic. Have a person check quoted statements, financial figures and names by ear. Avoid automatic translation directly into publishable subtitles without a bilingual review; language expansion changes line length and reading speed. For dubbed material, align the new audio, not the old English timeline.
The role of diarization
Diarization adds speaker turns when needed, but captions should identify a speaker only when that identity is supported. If the script already includes an actor list, pair those approved labels with the actual audio timing. Never present an acoustically inferred identity as a verified guest name. Make non-speech information such as relevant sound effects available where accessibility guidelines and the intended audience call for it.
Technical acceptance gate
Pass each file through encoding and time-range validation, compare total subtitle duration to media duration, search for overlaps and visually spot-check on mobile. Archive the checked script, audio checksum, model version and editor’s adjustments. A machine-generated VTT file is not accessibility-ready until it is readable at normal viewing speed and accurate enough to support understanding.
Worked production example
Suppose a six-minute voiceover has a trustworthy script but the speaker slows down during a dramatic passage. Dividing the total duration by character count will drift; forced alignment maps the actual spoken recording to known words instead. After alignment, group phrases so a caption does not flash a single word for a few milliseconds. A dramatic pause can be left empty or represented editorially where meaningful rather than attaching all surrounding words to an inflated cue. Playback speed and viewer reading time remain human design choices.
Practical acceptance gate
A technical verifier should reject negative cue durations, out-of-order timestamps, accidentally overlapping cues and text that differs from the approved narration. Test both WebVTT and SRT in the actual platform that will publish them. For a translated track, create new target-language timing from that final dub. The checker should record the corrected timestamps and editor name so a later audio rerender triggers a caption rebuild rather than leaving stale subtitles on a new soundtrack.
Region, plan and model-version notes
Documentation checked on 2026-10-25. Exact applicability: availability varies by account/endpoint and jurisdiction. Version: workflow (not a model). Plan: not independently verified for a specific plan. Monetary prices and quotas are not asserted unless separately verified; check the canonical AI 10 pricing ledger before purchase or publication of a tariff.
Sources & Further Reading
- Eleven Align — official documentation (checked 2026-10-25)
- Whisperx — official documentation (checked 2026-10-25)
- Open Whisper — official documentation (checked 2026-10-25)