9th Annual Conference of the International Speech Communication Association

Brisbane, Australia
September 22-26, 2008

Effects of Allophones on the Performance of Korean Speech Recognition

Hyejin Hong, Sunhee Kim, Minhwa Chung

Seoul National University, Korea

This paper investigates the effects of allophones on the performance of Korean speech recognition systems. Along with a baseline phone-like unit (PLU) set consisting of phonemes, 31 allophone-based PLU sets are designed by systematically considering 5 major Korean allophonic constraints which can describe all the PLU sets currently used for Korean speech recognition systems. Experiments for phone, word, and continuous speech recognition are performed using the proposed PLU sets. The results show that the allophone-based PLU sets improve recognition performance compared to using a baseline phoneme-based PLU set. The performance improvement is clearly evident in phone recognition for isolated speech and in isolated word and continuous speech recognition using context independent units. As predicted, the performance improvement is less evident when context dependent (CD) units are used in the experiments, since the allophonic information is internalized in the CD units. Finally, the constraint Voicing-Lax is observed as playing a positive role compared to other constraints that are only partly influential.

Full Paper

Bibliographic reference.  Hong, Hyejin / Kim, Sunhee / Chung, Minhwa (2008): "Effects of allophones on the performance of Korean speech recognition", In INTERSPEECH-2008, 2410-2413.