Looking at the VGGish paper itself, I see they use spectrograms as inputs, they show results where they can identify instrument types. I'm not too sure how specific embeddings from these models can be. Do we know if spectrograms can differentiate between two people's voice?