A study on the musical source separation method using NMF and musical cues

Our research is to develop a signal processing process for separating musical audio signals into streams of individual sound sources. In this paper, we present a method that uses NMF method and some musical cues for musical source separation. The conventional separation method based on NMF method classify the separated note events into each stream 'manually', so the conventional methods are difficult to use in the real engineering. However, our method performs 'automatically' the classification process different from the conventional methods. The proposed process consists of the separation step and the reconstruction step. In the separation step, the audio stream is divided into "musical events," groups of the notes which have same frequency structures. The separation method is based on the Wang's method, which used Non-negative Matrix Factorization (NMF) method. In the reconstruction step, the divided note events are automatically grouped into streams of the individual sound sources using the musical cues. The proposed musical cues consist of the timbre features, the temporal features, and the pitch components. The proposed separation system is evaluated with some musical signals which contain the multi musical sources. The evaluation shows the proposed method can perform 'automatically' the separation using the proposed cues and have the same performance as manual classification.