The problem this solves
A transcript tells you what was said. A subtitle file needs to know when. Between those two facts sits an afternoon of scrubbing through a waveform, placing an in-point, listening, nudging it two frames, and moving on to the next line — repeated for every cue in the file.
The manual version is not just slow, it is inconsistent. The first fifty cues get careful attention and the last two hundred get whatever rhythm the editor fell into. Viewers notice this even when they could not name it: the subtitles feel loose at the end of a long video, because they are.
Aligning the transcript to the audio produces every timestamp from the same process, so the whole file is uniformly tight. It also makes re-timing free: when the edit changes, you re-run against the new audio instead of re-placing everything by hand.
What "word-accurate" does and does not mean
The aligner knows the start and end of every individual word, which is what lets it cut cues on phrase boundaries rather than on a fixed character count. It is not, and does not claim to be, frame-accurate to the millisecond on plosives — no aligner is, because the acoustic onset of a "p" is genuinely ambiguous across speakers and microphones.
In practice the error is well under a tenth of a second, which is roughly two frames at 25 fps and below the threshold where a viewer perceives lag. Where it matters is in precision editing work — matching a subtitle to a hard cut, or to a mouth shape in an insert shot — and there, exporting VTT and nudging the two or three cues that share a frame with the cut is faster than placing all of them.
- Word-level timestamps available in the export, not just cue-level
- Cue boundaries chosen to avoid splitting a phrase mid-clause
- Confidence score per line, so you know which lines are worth checking
- Re-run against new audio when the edit changes — the transcript does not need re-uploading
Preparing the transcript so the alignment holds
The single biggest factor in output quality is whether the text matches what was actually recorded. Three habits make the difference:
- Keep it to what was said. Stage directions, speaker labels like "HOST:", and parenthetical notes will be treated as words. Strip them first, or the aligner will try to find audio for text that was never spoken.
- Let it be one line per cue, or one paragraph — both work. What does not work is a hard-wrapped file where the author pressed Enter at 72 characters; those wraps become cue boundaries and you get cues that end mid-sentence.
- Numbers and abbreviations are read as written. "2024" will be matched against "twenty twenty-four" and usually succeeds; "Dr." against "doctor" usually succeeds too. When it does not, the line is flagged rather than mistimed.
When the transcript and the audio diverge
They always diverge a little. A speaker adds a word, skips a clause, repeats themselves for clarity, or reads the same line twice for a better take. The aligner treats these as local mismatches and routes around them — the surrounding lines stay correctly timed, and the divergent line loses confidence.
What it cannot do is invent audio. If a paragraph exists in the transcript but was cut from the recording, that paragraph has nothing to attach to. It comes back flagged, with the surrounding timings intact, and you delete it. That is a much better outcome than the timings silently collapsing for everything after the cut, which is what happens in aligners that assume the two inputs are in sync.
From flat text to timed cues
Note the first cue: it cuts before "approach" rather than after "first", because "the obvious approach" is a noun phrase and splitting inside it reads badly. The aligner had both options and chose the one that does not break the phrase.
Right, so we tried the obvious approach first and it did not work at all.
1 00:00:01,760 --> 00:00:04,900 Right, so we tried the obvious 2 00:00:04,900 --> 00:00:07,240 approach first and it did not 3 00:00:07,240 --> 00:00:08,880 work at all.
Versus placing cues by hand in an editor
This is the comparison that matters, because it is what most people are actually doing now. Transcription tools are the other alternative, and they solve a different problem — see the audio-to-SRT page for that one.
| The usual way | AlignSRT | |
|---|---|---|
| Time for a 60-minute file | 3–6 hours of scrubbing and nudging | Minutes, mostly upload and queue |
| Consistency across the file | Degrades — attention is not infinite | Every cue from the same process |
| When the edit changes | Re-place everything by hand | Re-run against the new audio |
| Cost of a re-timing pass | Another afternoon | One more job against the same script |
| What review looks like | Watch the whole thing, twice | Check the flagged lines |
Questions about this
How is this different from just transcribing the audio?
A transcriber produces text you then have to check. This takes text you have already checked and produces timings. If your text is right, the output text is right — there is nothing to proofread, only lines to spot-check where confidence is low.
Can it align a transcript that has speaker labels?
It will run, but the labels are treated as spoken words and will throw off the confidence on those lines. Strip "HOST:" / "GUEST:" prefixes before uploading and re-add them in your editor if you need them — SRT has no field for a speaker, so they are cosmetic anyway.
What if the recording has more than one speaker?
It handles conversational audio fine as long as people are not talking over each other. Genuine overlap — both voices at once — is the hard case for any aligner. Split the channels and align each separately if that is what you have.
Do I get word-level timings, or only cue timings?
Both. SRT is cue-level by format, so the cue file is what you normally want. The word-level timestamps are available in the export for karaoke-style highlighting, animated caption burn-in, or anything that needs the word rather than the sentence.
Is 4 hours really the limit?
Yes, per file, with no surcharge for the length. Audiobook chapters and full lecture recordings are the reason the ceiling is that high — the alternative, splitting a chapter into four parts and stitching SRTs back together, invites off-by-a-frame errors at every seam.