Nonstationary Bayesian Independent Component Analysis

• Main Authors: Shih-husiung Li, 李仕雄
• Other Authors: Jen-tzung Chien
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/15612008396397801403

碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 97 === In an intelligent speech perception system, it is required to recover speech signals from the mixed signals where some unknown and independent sources are simultaneously acquired by the system microphones. As we known, the independent component analysis (ICA) is a popular approach for blind source separation (BSS) and is referred as an important issue in the fields of machine learning. Traditionally, the standard ICA assumes that the source signals are stationary. This assumption restricts the performance of ICA in real-world applications. Since the source signals may be moving or may be active or inactive as time goes on, we propose a nonstationary Bayesian ICA (NB-ICA) for dealing with the nonstationary blind source separation for an intelligent speech perception system. The proposed NB-ICA algorithm is based on the online Bayesian learning theory which identifies the source activity and estimates the number of source signals in real time. Also, comparing with the other nonstationary ICA algorithms, the computation cost of the proposed NB-ICA can be significantly reduced. In addition, the segmentation information of the source signals is important for the applications of intelligent speech perception such as the systems of speech recognition, speaker identification and speaker diarization. Accordingly, we incorporate a semi-Markov model for capturing the duration information for the estimated status of the source signals. The experimental results show that the proposed NB-ICA algorithm is efficient for the separation of nonstationary source signals in terms of signal-to-inference ratio and detection accuracy of signal segments.