Value error when trying to cut audio

[nanometer34688]

I have found an issue while using the GRID dataset. The audio/video files range in length. Using the audio_preprocessing script, the option 'max_wav_length' seems to fail when wanting to set the desired length of the audio.

So for example if i run the option with 42000 as the max wave length, I get this error:

audio_samples[i, n_fft//2: len(samples) + n_fft//2] = samples ValueError: could not broadcast input array from shape (42240) into shape (42000)

I'm assuming that it should cut the audio up to 42000 samples? Am i correct in thinking this?

[dr-pato]

Hi, actually not. If you read the help of 'audio_processing' subcommand you can find the description of '--max_audio_length' option:

Set this value to the maximum length (in samples with desidered sample rate) of single wav

You can easily modify the code to take shorter segments for each sample.

Giovanni

[nanometer34688]

That's the option I am having trouble with. If an audio file has a length longer than 42000, i get that error.

This is the command I run:

python3 av_speech_enhancement.py audio_preprocessing --data_dir zipped_data/TEST_SET --speaker_ids 2 8 --dest_dir . --audio_dir . -ml 42000

And this is the error i get:

audio_samples[i, n_fft//2: len(samples) + n_fft//2] = samples ValueError: could not broadcast input array from shape (42240) into shape (42000)

Is this something you have seen before?

[nanometer34688]

On the GRID Copus dataset, did you use the default value of 48000 for your max_audio_length option?

[nanometer34688]

@dr-pato I seem to have found the issue.

audio_features.py line 39 is:

audio_samples[i, n_fft//2: len(samples) + n_fft//2] = samples

But it causes an error as samples can be larger than audio samples.

The fix i made was as follows:

audio_samples[i, n_fft//2: len(samples) + n_fft//2] = samples[:max_audio_length]

Now my audio files have been cut and have now fixed this issue.

Thank you