Speech to text causing space leak in batch mode

[xjlin0]

On ARM MacOS Tahoe with Whisper CPP (large v3 turbo), when I select many long mp3 files, converting them will somehow take disk space more and more, eventually consume all space available until SubtitleEdit stops working. I wonder if all temp files generated for each file deleted in each iteration? Thanks for your work!

[cookesan]

This looks like a real temp-file lifetime issue in the batch path.

In the current SpeechToTextViewModel, each non-16 kHz WAV input is converted to a temp WAV under Path.GetTempPath() and added to _filesToDelete. Batch mode then calls StartNext(...) for the next job without clearing those files; DeleteTempFiles() is called on window close, but I don't see a per-item cleanup after each successful/failed batch item.

That means a long batch of long MP3s can legitimately grow temp storage until the batch/window finishes. On macOS, Path.GetTempPath() will typically be under the user's temp area, so the growth may not be inside the Subtitle Edit data folder.

A useful confirmation would be to watch the temp directory during the batch and look for GUID-named .wav files increasing. If that is what you see, the fix should be to delete the just-finished job's temp WAVs after StartNext(...) has written the subtitle, while still keeping cleanup-on-close as a fallback for canceled runs.

[niksedk]

Thx :)
Should be fixed in next beta.

[cookesan]

Fantastic!

[xjlin0]

Awesome team! Thanks!