Improved vocal isolation from varying repeating structures for musical signals

Abstract

A musical signal is the combination of a repeating background signal (music produced by musical instruments such as guitar, tabla, sitar, etc.) and a non-repeating, variable foreground signal, such as vocals or the human voice. By utilising this property of the audio mixture, a technique for isolating the non-repeating vocal or foreground from the repeating background in the audio mixture is enhanced so that it functions well even when the repeating background varies over time. The ability to quickly separate a song into its musical and vocal components would be useful for a variety of applications, including instrument and vocalist recognition, pitch and melody extraction, audio post-processing, and karaoke game development. This technique for extracting repeating patterns is effective when the repeating background is stable and the amount of distortion is minimal. Due to the error in the estimation of the repeating period, the efficacy of this technique deteriorates. Therefore, a more accurate method for estimating the period of repetition is required. Consequently, this estimation of the repeating period is improved with the aid of an independent component analysis algorithm and the repeating pattern extraction method, and its performance can be further enhanced by computing the mean segment before the median computation in the repeating segment modelling stage. Various analysis methods, including spectrogram and waveform analysis, energy analysis, source-to-distortion ratio analysis, false rejection rate analysis, and time complexity analysis are used in validating the improvements brought by the proposed approach compared to the existing algorithms.