I am trying to get subtitles for a long audio file, using Medium. It's the audio from a compilation from a spanish-language TV series.

My command was this:

pwcpp \
  --model "medium-q8_0" \
  --language "es" \
  --temperature 0.1 \
  --print_realtime true \
  --output-srt \
  ./input.wav

I ran it, saw it was doing great, so I left for a few hours. I came back to find it went crazy at around the 2 hour markk:

[02:03:58.320 --> 02:03:59.320]   ¿Qué pasa?
[02:03:59.320 --> 02:04:00.320]   ¿Qué pasa?
[02:04:00.320 --> 02:04:01.320]   ¿Qué pasa?
[02:04:01.320 --> 02:04:02.320]   ¿Qué pasa?
Progress:  37%
[02:04:02.320 --> 02:04:14.320]   ¿Qué pasa?
Progress:  38%
[02:04:14.320 --> 02:04:39.320]   ¿Qué pasa?
Progress:  38%

...repeat ad nauseum for 2.5 hours

[04:34:56.320 --> 04:35:11.320]   ¿Qué pasa?
Progress:  84%
[04:35:11.320 --> 04:35:26.320]   ¿Qué pasa?
Progress:  84%

I had previously tried another whisper.cpp-based transcription tool (Memo, using the non-quantized Medium model) and the same thing happened after 30min, with a different repeating line.

Is whisper.cpp inherently unreliable for long-running tasks? And could pwcpp provide a way to mitigate this? For example:

• I didn't even get a partial srt file when I killed pwcpp. Always write the srt even if incomplete.
• Provide the user a way to resume. Let us specify the timestamp to start at, and a file to resume from --begin-at-timestamp "02:04:39.320 --use-existing-srt ./previousoutput.srt"
  (or at least just --begin-at-timestamp. I could manually combine the two srt files.)