I'm super interested in their choice of features here; I never expected they'd be using the image of the spectrogram fed into a vision model like a CNN rather t

I'm super interested in their choice of features here; I never expected they'd be using the image of the spectrogram fed into a vision model like a CNN rather than the audio signal itself. My hazy memory from school was that speech recognition typically used MFCCs (Mel-frequency cepstral coefficients), which were some sort of manipulation of the frequencies. I don't remember how that really works but it seems more straightforward than turning an audio problem into a visual problem. I'm not critiquing, as I've used the app and it's excellent, just wondering if anyone could explain more about how feature engineering works in audio and whether this approach by the Merlin team is a standard one or not.