Separation of Voice and Music from Myanmar Songs using FFT-based FIR Filter and Repeating Pattern Extraction Technique (REPET) Method

Abstract

The audio signal represents sounds that may contain speech, song and all types of sounds. The audio signal processing provides many research areas such as speech recognition, audio events recognition, music information retrieval (MIR). Separation of singing vocal sound and music from the song signal is one of the important parts of music information retrieval (MIR). The singing vocal sound and music parts can usually be separated from the repetitive features of music. This paper performs to separate the voice of singer and music pieces of Myanmar songs. In this proposed work, the original song signal is first filtered with the FFT- based FIR filter to equalize the speech signal and music signal of the original input mixture song. After that, the repeating pattern extraction technique (REPET) method is utilized to extract the song’s background music. The separating performance of this algorithm is presented with the values of the signal to distortion ratio (SDR), signal to interference ratio (SIR), signal to artifacts ratio (SAR) that evaluated with the blind source separation evaluation (BSSEVL 3.0) Matlab tool.