自組織映射圖應用於聽覺場景式語音分離

Abstract

過去十年間，聽覺感知的一些細部的特性被大量的應用在語音處理的演算法中以提升效能。例如：在語音分離的領域中，使用多個麥克風的演算法如獨立成份分析(Independent Component Analysis, ICA)經常被使用而且有令人滿意的成果。然而，人類並只需要單耳便能將混合的聲音分開。本論文中，我們設計一個基於聽覺感知模型的單耳語音分離系統。我們從此模型中取出不同在時域-頻域上的一些使用於單耳語音分離系統的線索，之後，利用自組織映射圖來模擬神經元將混合的語音分組和歸類成分開的語音。最後，我們將比較分開語音和原來語音來顯示出本系統的效能。

During the past decade, detailed characteristics of auditory perception have been largely incorporated into speech processing algorithms to enhance their performance. For example, in the field of sound segregation, algorithms good for the condition of multiple microphones, such as independent component analysis (ICA), are often used and show satisfactory performance. However, the truth is human has no problems in segregating mixed sounds with only one ear. In this thesis, we design such a monaural speech segregation system based on an auditory perceptual model. Various spectral-temporal cues extracted from the model are used for monaural speech segregation. Then, a self-organizing feature map neural network is utilized to mimic the neural function in segregating and clustering a mixed sound into separated sounds. At the end, we demonstrate our system’s performance by comparing the separated sound with original sound.