A Novel Independent Component Analysis for Noisy Speech Recognition

• Main Authors: Chang-Kai Chao, 趙昶凱
• Other Authors: Jen-Tzung Chien
• Format: Others
• Language: zh-TW
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/49923643104009055088

碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 95 === Independent component analysis (ICA) is a widely accepted mechanism in solving blind source separation (BSS) problem. In this study, we develop a new ICA approach for unsupervised learning and apply it for hidden Markov model (HMM) clustering and noisy speech recognition. The underlying concept of proposed ICA algorithm is to de-mix the HMM mean vectors and identify the corresponding mixture sources prior to HMM clustering. These independent sources represent the specific noise conditions embedded in speech signals or features. We focus on presenting a general unsupervised learning algorithm based on a new ICA objective function. We will apply this algorithm for BSS and different problems in speech recognition. In ICA procedure, we follow up a predefined objective function measuring the dependence among feature vectors and derive an optimal demixing matrix, which can minimize the measure of dependence or mutual information. The basic assumptions of ICA include the source signals being mutually independent and having non-Gaussian distribution. We are using the Jensen’s inequality to derive a new metric of dependence measure. The parametric and nonparametric ICA approaches are developed. The generalized Gaussian model is used to characterize the non-Gaussianity of an acoustic random vector. We exploit a parametric ICA using generalized Gaussian distribution and also a nonparametric ICA using the Parzen window based distribution. We evaluate the efficiency and effectiveness of the proposed objective function in finding ICA demixing matrix compared to the existing objection functions including maximum likelihood, minimum mutual information and maximum non-entropy, etc. In the experiments, we investigate the performance of BSS and noisy speech recognition. We are using this ICA method for HMM clustering and evaluating speech recognition performance on AURORA2 noisy speech database. The preliminary results show that the proposed ICA achieves faster convergence property and higher recognition rate.