Can FFT be used to cluster sound waves based on their similarity?

Question

I am new to this so apologies if the question appears trivial.

Say we have n sound files and we want to cluster them to identify which ones are more similar. I roughly remember that FFT can be used to decompose the sound and analyse the frequencies of the different components.

If two sound files have exactly the same frequencies, does this also mean that they have very similar sound patterns?

Or am I going down the wrong way?

Answer 1

Sure it's a reasonable idea. Using spectrograms, which is a similar idea, is a somewhat popular method in speech processing and machine learning. (E.g. see here). One can use such features for clustering as well of course.

Maybe an interesting paper for you is Deep clustering: Discriminative embeddings for segmentation and separation by Hershey et al.

If two sound files have exactly the same frequencies, does this also mean that they have very similar sound patterns?

Um, well, mathematically, if they have exactly the same frequencies, then they have identical sound. This is because the Fourier transform is invertible :) The DFT is likewise bijective.