An Exploration of Mood Classification in the Million Songs Dataset

Publication Type:

Conference Paper


The 12th Sound and Music Computing Conference, Music Technology Research Group, Department of Computer Science, Maynooth University#, Maynooth, Co. Kildare, Ireland (2015)





Constraints, Harmonic/Percussive separation, Latency, Non-negative Matrix Factorization, Signal to Distortion Ratio (SDR), Signal to Interference Ratio (SIR), Smoothness, Sound source separation, Sparseness Online


The separation of percussive sounds fromharmonic sounds in audio recordings remains a challenging task since it has received much attention over the last decade. In a previous work, we described amethod to separate harmonic and percussive sounds based on a constrained Non-negative Matrix Factorization (NMF) approach. The approach distinguishes between percussive and harmonic bases integrating
percussive and harmonic sound features, such as smoothness and sparseness, into the decomposition process. In this paper, we propose an online version of our previous work. Instead of decomposing the whole mixture, the online proposal decomposes a set of segments of the mixture selected by a sliding temporal window. Both percussive and harmonic bases of the next segment are initialized using the bases obtained in the decomposition of the previous segment. Results show that an online proposal can provide
satisfactory separation performance but the sound quality of the separated signals depends inversely on the computation time of the system.
SMC2015_submission_3.pdf236.22 KB
SMC paper: