`segments[-1].end` answers a different question from the one the trimmer was asking. It reports where the audio stops being speech. The code needed where the script stops being spoken. In clip base_11 those two timestamps are 1.60 seconds apart: the script ends at 3.50s, the extra segment runs to 5.10s, and 5.10s is the end of the file [12][17].
So the call trimmed a 5.10-second clip down to 5.10 seconds, plus 40ms of padding it had no room for [11][13]. The one thing it deleted reliably was trailing silence that was harming nobody. The transcript comparison then ran again, failed again, and the clip was excluded [13].
The second reproduction makes the geometry plainer. The script ends at 2.50s, the hallucination starts at 3.98s and runs to 5.84s [14]. That is 3.34 seconds of audio past the script with a 1.48-second silence sitting in the middle of it [18]. A gap that size is cheap to detect, but the trimmer never examined gaps; it had already committed to an array index.
When the transcriber returns one segment, `segments[-1]` is `segments[0]`, and cutting there removes trailing silence harmlessly. The author notes the intent was clear and unit tests would pass [15]. That is the clean path, and it was the only path the function was ever exercised on.
Detection came from the batch, not from the code. Twenty and eighteen rejections against a zero-to-ten band across the other voices [4]; on a 65-sentence base corpus at one clip per sentence, that is 31% and 28% against a worst healthy case of 15%, about double [5][19]. Ten sound voices are what made two bad ones legible [20]. A single-voice run would have produced a number nobody questioned.
Two conditions have to hold before this is your bug. Your TTS has to be deterministic for a fixed caption and seed, which is what turned an intermittent artifact into the same failure on every rerun [9]. And your transcriber has to split the tail into its own segment [12]. If yours folds the hallucination into a single segment, `segments[-1].end` is the correct expression and this failure mode never appears. Note also that with determinism in place, only one rung of the three-level ladder could change the audio at all: re-rolling `num_steps` with caption and seed pinned [10].
The mora correlation is suggestive rather than established. The five named rejections run 14 to 17 mora against a corpus median of 21 and a minimum of 13 [5][6], so all five sit 4 to 7 below the median and the shortest is one mora off the floor [21]. Five samples are not a curve across 65 sentences. If length is the driver, rejection rate should fall as mora count rises, and that is a plot the pipeline already has the data to draw.
The boundary the trimmer needed was in the payload the whole time: the end of the last segment whose text still matches the script, 3.50s in the first clip and 2.50s in the second [12][14]. The narrow lesson is about fixtures. A fallback that indexes into output produced by the very failure it handles has to be tested against that output, and a two-segment transcript is four lines of JSON.
