tscizzle
commented
2 years ago I'll add that confidences (say, per phone boundary) would be awesome!
Even in broad strokes, it'd be really helpful. For example, mumbled words which almost aren't intelligible in the audio, sometimes Gentle doesn't want to outright say "not-found-in-audio" and tries to force some phones where they don't quite belong. Maybe Gentle is on the fence about if it might as well try and place those phones instead of give up and say "not-found-in-audio", but if it knows that it's on the fence, it could indicate as such with a "confidence" between 0 and 1.
Not sure what the best one would be... This would helpful to get higher quality results when doing supercut-like experiments with large datasets.