5.3.0 Probem with speach to text Portugese and Italian (so far)

[episode10]

Working through Engines and models and keep getting this problem where only the first line or first few lines are created then the rest of the output is the same line repeated to the end, on any combination of engine and models that work together
.
Experiencing this on both Portuguese and Italian Videos
.
Windows 10 and SE 5.3.0
.
It's making me miss the older version as it worked!

[Columbo]

Claude says:

This is almost certainly Whisper’s well known repetition loop, not a Subtitle Edit bug. Once the model emits a line, it feeds its own previous output back in as context for the next window, and if it latches onto that text it keeps predicting the same line until the end of the file. Newer engine builds and models (large-v3 especially) are more prone to it, which is likely why an older setup worked for you.

Things that stop it, entered in the Advanced parameters box of SE’s Audio to Text (Whisper) dialog:

whisper.cpp / CPP engines
--max-context 0
This disables feeding previous text back as a prompt and is usually the single most effective fix.

Purfview’s Faster-Whisper-XXL / faster-whisper engines
--condition_on_previous_text False
Same effect as above. Adding --vad_filter True also helps, since loops often start on silence or music. XXL also accepts --hallucination_silence_threshold 2.

OpenAI / Const-me engines
--condition_on_previous_text False (OpenAI). Const-me has no equivalent setting and is generally the worst for this, so I’d avoid it.

Model choice
If you’re using large-v3, switch to large-v2 or large-v3-turbo. large-v3 is noticeably worse for repetition on non-English audio.

Set the language explicitly
Choose Portuguese or Italian in the dialog rather than auto-detect. Auto-detect on an ambiguous opening segment can throw the model off and trigger a loop.

Disabling previous-text conditioning may make punctuation and capitalization slightly less consistent between lines, but it removes the runaway repetition.

[NovaHunter47]

That repeated-line tail is usually a decoder loop triggered by silence/music or a bad chunk boundary, not necessarily the language model itself. Try forcing the language explicitly, disabling VAD/chunking, and testing a short 30-second clip with a non-quantized model; if short clips are clean but long files loop, its likely a 5.3.0 segmentation issue worth narrowing down.