Chinese-English mixed-lingual keyword spotting

Abstract

Base on our former experience in the "ITRI 104 Auto Attendant System" of using keyword spotting for Mandarin speech recognition (W.-C. Shieh et al., CCL Technical Journal, vol. 96), a Chinese-English mixed-lingual keyword spotting system, which caters for the Taiwanese speaking style, is presented. Detailed descriptions and discussions for developing the mixed-lingual auto attendant system are included, especially for solving different scoring scales in the decoding phase and the re-scoring phase for the two languages. In the decoding phase, we propose a bias-compensation method to make up the score-gap in the likelihood calculation of using Chinese and English acoustic models. To select the most probable result from the recognized hypotheses, a method is also presented of normalizing the combination scores when using different scoring mechanisms in the re-scoring phase.