A Method to Separate Musical Percussive Sounds using Chroma Spectral Flatness

摘要

This paper presents an unsupervised Non-Negative Matrix Factorization (NMF) approach to extract percussive sounds from monaural music signals. Due to unconstrained NMF cannot discriminate between percussive, harmonic or singing-voice components in the decomposition process, we propose a novel method to extract percussive sounds based on the anisotropic smoothness of percussive chroma. Thus, percussive sounds can be discriminate because chroma from percussive sounds clearly draws lines along the chroma. Under a NMF framework, a time-domain signal related to a component is labelled as percussive is the energy distribution of its chroma is approximately flat. This proposal does not require information about the number of active sound sources neither prior knowledge about the instruments nor supervised training to classify the bases. Real-world audio mixtures composed of Harmonic/Percussive and Harmonic/Percussive/Singing-voice sounds were evaluated. Experimental results showed that the proposal was effective compared to state-of-the-art methods. An interesting advantage of the proposal is that it can remove most of the singing-voice components from the extracted percussive signals.