Improvements implemented in the audio processing module #2390

mrfelpa · 2024-10-15T18:06:21Z

The relative import causing errors when the script was executed directly, so I changed from .utils import exact_div to from utils import exact_div. I also implemented a function (get_hann_window) to avoid repeated Hann window calculations.

Nisarg236 · 2024-10-16T03:29:24Z

whisper/audio.py

+
+   log_spec = torch.clamp(mel_spec, min=1e-10).log10()
+
+   log_spec_normalized = (log_spec + 4.0) / 4.0


maybe combine both?
log_spec_normalized = (torch.clamp(mel_spec, min=1e-10).log10() + 4.0) / 4.0

@Nisarg236 I did it separately to try to keep the code clearer, but yes, it is possible to combine the two.

Improvements implemented in the audio processing module

81ef7f3

Nisarg236 reviewed Oct 16, 2024

View reviewed changes

Provide feedback


		log_spec = torch.clamp(mel_spec, min=1e-10).log10()

		log_spec_normalized = (log_spec + 4.0) / 4.0