7th International Conference on Spoken Language Processing

September 16-20, 2002
Denver, Colorado, USA

Application of Over-Complete Blind Source Separation for Robust Automatic Speech Recognition

Shubha Kadambe

HRL Laboratories, USA

Spoken dialogue based information retrieval systems that are used in mobile environments are becoming popular. However, mobile environment is dynamically changing and there exists many interfering signals. These two effects result in degradation in automatic speech recognition (ASR) accuracy and hence, degradation in performance of spoken dialogue based information retrieval systems. One way to improve the speech recognition accuracy is to separate the intended speech signal from the interference signals and use the enhanced speech signals in recognition. In this paper, we describe a technique that we applied for speech signal enhancement. We also provide the relative improvement in recognition accuracy that we obtained by using such enhanced speech signals in an ASR system. For speech signal enhancement, we apply the Over-complete Blind Source Separation (OCBSS) technique that we developed. For ASR, a continuous speech recognizer was used. In this paper, we also compare the recognition accuracy results of another BSS technique that is based on Independent Component Analysis (ICA) - JADE-ICA with OCBSS. The results indicate that as the complexity of signal separation problem increases i.e., close to real scenarios, the OCBSS provides about 30% better relative improvement in recognition accuracy as compared to JADE-ICA.


Full Paper

Bibliographic reference.  Kadambe, Shubha (2002): "Application of over-complete blind source separation for robust automatic speech recognition", In ICSLP-2002, 805-808.