Harmonic and percussive source separation using a convolutional auto encoder

Real world audio signals are generally a mixture of harmonic and percussive sounds. In this paper, we present a novel method for separating the harmonic and percussive audio signals from an audio mixture. Proposed method involves the use of a convolutional auto-encoder on a magnitude of the spectrogram to separate the harmonic and percussive signals. This network structure enables automatic high-level feature learning and spectral domain audio decomposition. An evaluation was performed using professionally produced music recording. Consequently, we confirm that the proposed method provides superior separation performance compared to conventional methods.