Why forced alignment beats AI timestamps

Speech-to-text models like Whisper are astonishing at writing down what was said, and mediocre at telling you when. That's not a bug to be patched - it's a consequence of how they work. Forced alignment is a different tool that answers the "when" question properly.

1. Where AI timestamps come from

Transcription models process audio in chunks and emit text tokens. Word timestamps are reconstructed afterwards from the model's internal attention patterns - essentially, an educated guess about which moment of audio the model was "looking at" while writing each word. It was never trained to be a stopwatch, and it shows: word boundaries land early or late, short words get folded into neighbours, and silences confuse it.

2. Why the drift is worst where it matters

A subtitle that appears 300 ms late is something viewers feel before they can name it. The drift is worst exactly where captions need to be sharpest: fast speech, sentence boundaries, and pauses. For word-by-word highlighted captions it's fatal - a highlight that lands on the wrong word looks broken, not stylish.

3. What forced alignment does instead

A forced aligner is given two things: the audio and the final text. It breaks each word into phonemes - the individual speech sounds - and finds the acoustic moment each one is actually pronounced. Because the text is known in advance, our system spends all of its capacity on timing, not on figuring out what was said. The result is per-word start and end times measured from the audio itself.

4. Why the text review matters

Alignment is only as good as the text it aligns. If the transcript says "the Sigma lens" but the audio says "this Sigma lens", the aligner has to force a mismatch. That's why our flow puts a review step before alignment: fix any misheard words, or skip transcription entirely by uploading your narration script. Better text in, frame-accurate timing out.