I'm back after a long time with a problem that is causing me quite a few headaches. I have two audio tracks: The “optimal” track is a recording of a hydraulic press, repetitive and constant. The “anomalous” track is also a recording of the press, but with music. The goal is “simple”: to filter the optimal track from the anomalous track so that only the “music” can be heard. What I have already tried without success (in Python): Autoencoder on MFCC with reconstruction error Spectral subtraction (post MFCC extraction and alignment) Direct spectral subtraction (in the time domain) Various filters, e.g. Wiener Spleeter and Demucs (source separation) Audio Source Separation Using Non-Negative Matrix Factorization (NMF) None of these have resulted in clean filtering; in the resulting track, I can still “hear” frequencies from the base track. Perhaps one of you can enlighten me. Continue reading...
