> image representations of audio are horrible representations of it The spectrogram is just a series of FFTs taken over time; encoding it as a bitmap doesn't r

> image representations of audio are horrible representations of it The spectrogram is just a series of FFTs taken over time; encoding it as a bitmap doesn't really change this, aside from precision issues. Any other representation of the audio is derived from either the original time-domain signal or the FFT. Indeed, humans can't reliably map raw waveforms or spectrograms to intuitive musical phenomena. But a CNN should be able to derive meaningful features from these basic representations on its own.