Question for commons.slice_segments(wav_padded, ids_slice * 480, wav_seglen) about freevc HOT 4 CLOSED

This may be caused by data preprocessing, the audio length doesn't match spectrogram length.
librosa.effects.trim can have different trimed segment with different sampling rates. For example, wav, _ = librosa.effects.trim(<24k_wav>, top_db=20) might trim out the first 0.3 seconds, while wav, _ = librosa.effects.trim(<22k_wav>, top_db=20) might trim out the first 0.33 seconds.
The 24k wav and 16k spectrogram and 16k wavlm feature should be from the same segment.
And btw I just found that I forgot librosa loads 22k wav by default, the original downsampling code upsamples 22k wav to 24k wav. Just fixed it.

from freevc.

Thanks for your kindly answering.
Then, if there is no leading and trailing silence, is there no difference in length between 16k spectrogram and 24k wav?

from freevc.

If you didn't trim both 16k wav (where the 16k spectrogram come from) and 24k wav, then yes.
librosa.effects.trim might have the possibility to trim a little even if there is no apparent leading and trailing silence.

from freevc.

Thanks very much. After matching the length of 16k and 24k it works fine.

from freevc.