7th International Conference on Spoken Language Processing

September 16-20, 2002
Denver, Colorado, USA

Low Complexity Mandarin Speaker-Independent Isolated Word Recognition

Xia Wang, Juha Iso-Sipilš

Nokia Research Center, China

This paper addresses the problem of tone recognition in Mandarin Chinese language. The tone, or pitch contour, has a lexical meaning in Mandarin language. This means that changing the pitch contour of a syllable results in totally different meaning for the word. This is especially problematic in speech recognition applications where context does not allow the unambiguous resolution of the meaning. In order to cope with this, we have developed methods that allow us to recognize Mandarin isolated words that differ only by the pitch contour. Low complexity, both in terms of memory and computational complexity, has been another driver for this work. The target platform is an embedded system with limited memory and computation resources. Two new schemes are proposed with significantly lower complexity compared to the baseline system. High recognition accuracy of almost 90% has been retained despite more than 60% reduction in complexity.


Full Paper

Bibliographic reference.  Wang, Xia / Iso-Sipilš, Juha (2002): "Low complexity Mandarin speaker-independent isolated word recognition", In ICSLP-2002, 1589-1592.