Loading page
About this tool
Compare timed transcript cues with local audio activity and optionally diff them against a reference timed transcript.
The Audio-Transcript Alignment Checker measures your audio in 50 ms RMS windows, marks every window above -40 dBFS as speech activity, and reports what percentage of that activity falls inside an SRT or WebVTT cue — plus the specific stretches of audio that no cue covers. Paste a second transcript and it also pairs cues by time overlap and flags text differences using a Dice token-similarity score below 0.85 and boundary shifts beyond 500 ms. It is for captioners, QC reviewers and accessibility teams who need to find the missing or drifted cue rather than proofread the whole file.
Open Audio-Transcript Alignment Checker on AltFTool — it loads instantly in your browser.
Press 'Choose an audio file' (maximum 30 MB, decoded maximum 10 minutes) and paste your Candidate timed transcript as SRT or WebVTT, adding a Reference timed transcript if you want a diff.
Set Audio activity threshold (Sensitive -50 dBFS through Strong audio only -20 dBFS) and Cue timing tolerance (250 ms to 2 seconds), then press 'Run alignment screen'.
Read the percentage of active windows overlapping a cue under 'Playback and coverage' with the uncovered ranges and findings, then press 'Export counts and timings' to save audio-transcript-alignment-screen.json.
Uncovered speech is merged into contiguous ranges and reported with its peak dB and in/out times, so you get a worklist of timestamps instead of a single score.
Cues are paired by greatest time overlap first, so a renumbered or re-split file still matches up, and only then is the text compared — a rename does not read as a rewrite.
Findings are typed as reference-unmatched, candidate-extra, text-difference or timing-shift and counted individually, so a missing cue is never confused with a mistyped one.
Two are adjustable in this tool. Activity threshold (default -40 dBFS, presets -50/-40/-30/-20) decides what counts as speech; timing tolerance (default 500 ms, presets 250 ms/500 ms/1 s/2 s) is how far a cue boundary may drift before it is flagged. Three more settings — window size (50 ms), minimum uncovered length (250 ms) and text similarity threshold (0.85) — are fixed at their defaults for this hosted tool; the underlying library supports tuning them across a wider range, but this page does not expose controls for them.
As a Dice coefficient over word tokens: twice the number of shared tokens divided by the combined token count of both cues, giving 1.0 for identical wording and 0 for no shared words. Comparison is case-insensitive and ignores punctuation and markup, so `Hello, world!` and `hello world` score 1.0.
SRT and WebVTT blocks with `-->` timing lines in HH:MM:SS,mmm or MM:SS.mmm form, up to 1,000,000 characters and 5,000 cues; WEBVTT, NOTE, STYLE and REGION headers are skipped and tags and entities are stripped before comparison. Audio is limited to 30 MB, 10 minutes, mono or stereo, up to 96 kHz.
No, and the exported report says so explicitly. No speech recognition is performed, so the tool never knows what was actually said — it only compares energy in the audio against cue timings, and compares your two transcripts against each other. Semantic accuracy and WCAG conformance still need a human reviewer.