Is it a good practice to pad signal before feature extraction?

Question

Is padding, before feature extraction with VGGish, a good practice?

Our padding technique is to find the longest signal (which is loaded .wav signal), and then, in every shorter signal, put zeros to the size of the longest one. We need to use it because one size of input data is desirable.

Perhaps there is any other techniques you recommend?

The difference between padding before and after the features extraction by accuracy is quite big - more than 20%. Using padding before extraction gives 97% accuracy.

I'd be glad to read your feedback, and explain me why that happens, and tell me if that kind of padding is correct action or is there a better solution.

Answer 1

Padding is a common practice both in image-processing (typically via CNNs) and in sequence-processing tasks (RNNs, Transformers).

For CNNs all the standard convolutional layers - Conv1D, Conv2D and Conv3D,- have the padding argument. The padding values can be valid or same for 2d and 3d convolutions. And extra causal type of padding is possible for 1d convolutions and the documentation refers to this paper: WaveNet: A Generative Model for Raw Audio - which sounds quite close to what you are interested in.

This animations might be useful to get a bit more intuition about the convolutions and strides/padding. The general consensus is that using same padding is advantageous for model performance - your network gets more information about the borders of your inputs (and deeper networks are possible).

For sequential models padding is even more important. Training samples are usually of unequal lengths, so you have to pad them with a special token (usually called [PAD] that gets encoded as 0). Here are some examples of this mentioned in tensorflow docs, huggingface.transformers or BERT tutorial.