A Singing Voice Separation Method from Persian Music Based on Pitch Detection Methods

Singing voice separation algorithms have many applications, such as recognizing the lyrics, identifying the singer, and retrieving the music data. Due to its complexity and special rules, the traditional music and the voice of Iranian singers have caused the lack of sufficient research in this field. However, in western music, due to the relatively simple structure of the sounds, it has made it possible for researchers can easily do segmentation based on the recognition of its kind, using the characteristics like the number of uttered words and the recognition of the pitches. In this research, an algorithm is presented for the separation of the singing voice from the music through the recognition of the pitches and based on the characteristic-based method. First, at the stage of recognizing the song's vocal by segmentation of the speech signal, the input is classified into vocal and non-vocal parts. Then music is recognized from the vocal using the pitch criterion and the characteristics of the ceptral coefficients. Finally, the stage of the separation from the pitch is done and is used for the classification. In this research, an automatic and effective method is presented. In the proposed method, based on the extracted characteristics and the frequency distances in segmented parts, the pitch of the singer's voice and the process of the separation from the music is done with acceptable accuracy. The quantitative results of this research demonstrate the success of the separation system by the presented method. The precision of the proposed method using a combination of characteristics and the Shuffled frog-leaping algorithm's metaheuristic algorithm is about 94%. The precision of the proposed method using the dimension reduction through MLP neural system demonstrates is 90%.