Audio source separation using signal processing and machine learning techniques

Abstract

newline Blind Source Separation (BSS) is a signal processing technique used to extract the sources from their observed mixture signals without knowing the nature of mixing process and information about the sources. BSS is used in a broad range of applications including wireless communication, biomedical signal analysis, image processing and text document analysis. This work aims to separate the sources from real-time recorded speech mixture signals using suitable algorithms in a typical office noisy environment. As a first step, FastICA algorithm that is capable to separate the sources from its overdetermined instantaneous mixture signals is analyzed. However the underdetermined convolutive mixture signals closely replicate the real-time recorded mixture signals. The FastICA algorithm involves some limitations in separating the real-time mixture signals. FastICA algorithm is not suitable for the separation of the real-time recorded mixture signals because of the presence of background noise sources and the convolutive mixing nature of acoustic sources. The convolutive mixture signals are converted into the frequency domain using Short Time Fourier Transform (STFT). The observed mixture signals become instantaneous mixture signals in the frequency domain. A complex FastICA algorithm is proposed to extract the sources from its overdetermined convolutive mixture signals overcoming the highlighted limitations suggested above. The Permutation problem occurred while processing the mixture signals in the frequency domain is solved by using a two-pass method with the correlation between the power ratios of the estimated source components. newline newline