Readaloud alignment drifts and an entire chapter ends up unaligned (transcribed with whisper large-v3)

My readaloud starts drifting partway into a chapter and never recovers. I transcribed with whisper-large-v3 on a CUDA image (latest-cuda-13.1.0, RTX 3060 Ti), the audiobook is a single long m4a, and the ebook and audiobook are the same edition. Raising the container memory from 2GB to 8GB made no difference. Is this an audio format problem, a CUDA version problem, or something else?

Resolution: Drift like this is almost always a transcription failure, not an alignment or audio-container problem. Check it first with the built-in report: open <your storyteller url>/api/v2/reports/<uuid>, where <uuid> is the id from the book's page URL /books/<uuid>. If a chapter shows something like "chapterSentenceCount": 414, "alignedSentenceCount": 0, nothing in that chapter matched and every sentence timing was interpolated, which is exactly what reads as drift. The usual cause is whisper large-v3 hallucinating — it gets stuck emitting the same sentence over and over for a stretch of audio, so there is nothing for the aligner to match. Re-transcribe the book with large-v3-turbo or tiny.en instead; large-v3 is noticeably more prone to this loop. Audio format is not the culprit: variable-bitrate m4a/mp3 input is transcoded to CBR before transcription, and a CUDA minor-version mismatch that does not error out immediately is harmless. Note that perfect sync is not achievable — most chapters miss one to three sentences, typically at the very start or very end, and any gap of two or more consecutive sentences has to be guessed.