Preprocessing audio samples for CNN training

[ctmittelstaedt]

Hi all,

I would like to preprocess the audio samples that I will use to train my CNN. I am following the “Preprocess audio samples” tutorial but am a bit confused about how the modifications to the preprocessor are actually applied. Here are the modifications that I would like to make to the samples in my training dataframe:

preprocessor = SpectrogramPreprocessor(sample_duration=0.3)
train_dataset = AudioFileDataset(train_df, preprocessor)
preprocessor.pipeline
preprocessor.pipeline.bandpass.set(min_f=700,max_f=6000)
preprocessor.pipeline.to_spec.params['overlap_fraction'] = 0.9
preprocessor.pipeline.to_spec.params

I believe the modifications are working because I get this output:
load_audio Action calling <bound method Audio.from_file o...
random_trim_audio Augmentation Action calling <function trim_aud...
trim_audio Action calling <function trim_audio at 0x00000...
to_spec Action calling <bound method Spectrogram.from_… etc.

Though I haven't confirmed this visually as I haven’t been able to get the show_tensor function to work.

I think I must be missing something because my recognizer performs identically with and without the preprocessor modifications, so I assume the samples being used during training are the ones processed by the default processor. I was wondering whether my modifications to the pipeline should automatically be applied to my training samples or whether I have to call on the newly preprocessed tensors during model training? If I do have to call on the new tensors, when would I do this? To my understanding, I can’t replace train_df with train_dataset during training as train_dataset isn’t a dataframe.

Any insight would be much appreciated!

Thanks,

Charlotte

[ctmittelstaedt]

Never mind! I'm getting different performance metrics now. Not sure what the issue was before :)