Rating Utterances
Status: Current Last updated: 2026-03-30 13:40 EDT
Rating values
| Rating | Key | Marker | Meaning |
|---|---|---|---|
| Good | 1 | [ok] | Bullet timing is correct — audio matches text |
| Early | 2 | [early] | Bullet starts before speech actually begins |
| Late | 3 | [late] | Bullet starts after speech actually begins |
| Wrong | 4 | [wrong] | Completely wrong — hearing different content |
| Skip | 5 | (unchanged) | Can’t tell or don’t want to rate — stays [?] |
What “Good” means
The bullet points to the audio location where the words in the utterance are actually spoken. A few milliseconds of imprecision is fine — “Good” means “I can hear the right words when I play from this bullet.”
What “Early” means
The bullet plays audio from before the utterance starts. You hear silence, the previous speaker’s words, or background noise before the current utterance’s words begin.
This is the most common issue with boundary-averaged bullets — the algorithm split the difference between two overlapping utterances, but split it in the wrong place.
What “Late” means
The bullet plays audio that starts partway through the utterance — you miss the first word or two. The speaker is already talking when playback begins.
What “Wrong” means
The bullet is wildly off — you hear a completely different part of the conversation, a different speaker, or content that doesn’t match the text at all. This typically happens when the alignment algorithm lost sync.
Adding notes
After pressing a rating key, the %xrev tier updates automatically. If
you want to add a note, edit the line directly in the editor:
%xrev: [early] starts about 200ms before speech, consistent with other INV lines
Notes don’t affect the rating — they’re free text after the marker. The harvesting script collects them for manual analysis.
Useful note patterns
- Timing offset: “early by ~300ms” or “late by about half a second”
- Speaker pattern: “all INV backchannels seem early”
- Overlap pattern: “this is where two speakers talk over each other”
- Algorithm feedback: “LIS shouldn’t have stripped this — timing was correct”
- Correction hint: “should start at roughly 12300ms, not 12095ms”