Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

A Recombination Strategy for Multi-Band Speech Recognition Based on Mutual Information Criterion

Shigeki Okawa (1), Takehiro Nakajima (2), Katsuhiko Shirai (2)

(1) Department of Network Science, Chiba Institute of Technology, Narashino, Japan
(2) Department of Information and Computer Science, Waseda University, Tokyo, Japan

This paper presents a recombination strategy for multi-band automatic speech recognition (MB-ASR). Several recent works have suggested that MB-ASR gives more accurate recognition, especially in noisy acoustic environments. The main issue in this study concerns the sub-band score recombination in MB-ASR framework. Intuitively, it seems very improbable that all sub-band features have the same amount of information for speech recognition. We therefore investigate to weight the contribution from each band at the recombination process by using a strategy derived from the information theory. The quantity of information is well determined by the mutual information between band features and target phoneme categories to be recognized. The experimental results show that the recognition accuracy improves for noisy speech by using three and six stream systems with the proposed approach.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Okawa, Shigeki / Nakajima, Takehiro / Shirai, Katsuhiko (1999): "A recombination strategy for multi-band speech recognition based on mutual information criterion", In EUROSPEECH'99, 603-606.