GitLab

I have added the argument normalize_wav to the following classes and functions (which all handle loading of audio from wav files),

• Waveform.from_wav
• MagSpectrogram.from_wav, PowerSpectrogram.from_wav, MelSpectrogram.from_wav, CQTSpectrogram.from_wav
• AudioLoader, AudioFrameLoader, AudioSelectionLoader
• create_database

If normalize_wav=True, the waveform will be normalized so it has zero (0) mean and unity (1) standard deviation. The default value is normalize_wav=False (ensures backward compatibility).

I have also expanded the audio representation parser to handle this new argument.

Unit tests have been created and are all passing.

Should be ready to merge. Please have a look, thanks

marked as a Work In Progress

changed title from Normalize waveforms to WIP : Normalize waveforms

changed the description

added 1 commit

• 25b36594 - normalize_wav passed on as part of audio representation

Compare with previous version

added 1 commit

• 0478e1d4 - parse_audio_representation can also handle normalize_wav argument

Compare with previous version

unmarked as a Work In Progress

changed title from WIP : Normalize waveforms to Normalize waveforms

changed the description

The Waveform and Spectrogram classes have this method 'normalize()', which they inherit from BaseAudio. But that puts the data into a 0-1 range.

So maybe it be would worth changing the implementation so that you can choose how to 'normalize'? I guess technically 'normalize' would mean 'put the data in a normal distribution with a given mean and a given std', but do you think it could be useful to keep the transformation that adjusts the range to. And, in both cases, would it be useful to have some parameters like the mean and std you want, or the range?

Good catch! I had forgotten about the normalize() method.

We could alter the implementation of the normalize() as you suggest, i.e.,

    def normalize(self, mean=0, std=1):
        """ Normalize the data array to specified mean and standard deviation
        """
        self.data = (self.data - np.mean(self.data)) / np.std(self.data)

And then have a separate method for adjusting the range,

    def set_range(self, range=(0,1)):
        """ Applies a linear transformation to the data array that puts the values
            within the specified range. 
        """
        x_min = self.min()
        x_max = self.max()
        self.data = (range[1] - range[0]) * (self.data - x_min) / (x_max - x_min) + range[0]

The latter method could also be named adjust_range instead of set_range ...

@fsfrazao , what do you think?

That looks great. And I'd go with adjust range.

Cool. I'll make the changes

added 1 commit

• be6f78ce - re-implemented normalize method and added adjust_range method in base_audio

Compare with previous version

Done! Should be ready to merge now.

added 106 commits

• be6f78ce...b48e2d16 - 105 commits from branch development
• 06d7e8e0 - Merge branch 'development' into 'norm_audio'

Compare with previous version

merged

mentioned in commit ea8d5b23