UC San Diego

The problem of separating mixed signals using multiple sensors with little to no information about the source signals is known as Blind Source Separation (BSS). Many embedded systems, such as cell phones, have an audio environment which routinely suffers from multiple simultaneous audio or noise sources interfering with the desired user. One approach for improving audio quality is to use BSS techniques such as Independent Vector Analysis (IVA), an extension to the more common Independent Component Analysis (ICA) approach, to remove interference or mitigate noise. However, these algorithms suffer from slow convergence rates, thus making it impractical to effectively separate sources in real time. This thesis explores Auxiliary Function Independent Vector Analysis (AuxIVA) with constraints on the frequency distribution of the audio components to improve convergence rates. AuxIVA has been shown to yield a faster convergence time and better results when separating audio sources compared to traditional IVA. By constraining input sources to be human speech, a harmonic frequency dependence model can be used to further improve convergence. We propose combining AuxIVA with a harmonic clique dependence model to achieve a more efficient algorithm, thus making a real time solution more viable. This thesis will demonstrate improved convergence rates and Signal to Interference Ratio (SIR) performance of the proposed technique relative to traditional IVA, AuxIVA, and AuxIVA with non-harmonic clique dependence.