Audio pre-processing for Machine Learning: Getting things right (opens in new tab)

It might be good to understand how changing the hop and window size affect the analysis so you're not blindly changing settings.

The trade off for window size is frequency resolution and time resolution. A bigger window gives you narrower bands, so more frequency resolution while giving you less temporal resolution where an onset of transient is significant in the analysis. Similarly, hop size will determine how 'leaky' the process is and how fine grained the windows will be. This can effect detecting quick peaks or changes while possibly smearing them across a few windows.

Hmmm, that makes sense to me for offline / not real time analysis, but I'd be interested to know how that would affect things in real time. I guess using some sort of DRC/limiter could be used to attempt to "normalize" the incoming audio

WAV is almost always lossless with PCM data. I'm not sure where you got the impression that "you don't see that in the wild too often". Depending on what kind of analysis you need to having your audio at 8k is going to deem any results useless. I would have it minimum 16k and aim for 44.1k in order to preserve the top end which is where a large quantity of useful information is. The reason most sets are recorded in 8khz is that they are running MFCC's which are quite stubborn and insensitive to the high end anyway with most enough information for machine learning existing in the bottom end. If you're doing music, or environmental sounds you really need to preserve the other frequency bands.