[abaichuan1017]

File transcription currently writes <stem>_resampled.wav into the working directory, overwriting an existing file or another conversion with the same name. It also leaves the converted file behind after playback.

Create a unique temporary WAV, close both PyAV containers on every path, remove failed conversions, and let the file-transcription client remove its owned temporary file after playback, including errors. Document the cleanup responsibility for direct resample callers.

Validation: 71 client/utility tests passed, including native PyAV conversion at 16 kHz, preserving an existing output, distinct conversion paths, conversion-failure cleanup, and cleanup after successful/failed playback. Three regression cases fail on the original implementation. git diff --check passed. No microphone, model, or server integration was needed for these file-lifecycle checks.

[boxerab]

thanks, this looks good. one gap: load_audio in whisper_live/transcriber/tensorrt_utils.py also calls resample and never removes the file. before this change it overwrote the same <stem>_resampled.wav each time, now every call leaves a new file in the temp dir. could you wrap the read in try/finally and os.remove(resampled_file) there too, like examples/manual_audio_chunking.py does?

optional: the completed flag could be except BaseException: unlink, raise instead.