應用轉換函式於具發音變異及台灣口音之英文語音辨識

Abstract

This paper presents an approach to acoustic model construction for the recognition of Taiwanese-accented English speech with pronunciation variation. This study fist analyzes the collected English speech database (EAT) to detect possible variations, and estimate the linear relation between the correct pronunciation and pronunciation variation. According to the relation, a transformation function is constructed to obtain other pronunciation variation models with similar pronunciation characteristics. This system can be applied to a speech recognition system to recognize the non-native speaker’s speech with pronunciation variation or accent. The models generated by the proposed method could achieve 50.91% word recognition accuracy for the speech data with pronunciation variation and accent. Compared to the system which uses the general acoustic models and obtains 23.26% word accuracy, the proposed method can achieve significant improvement. This approach can be applied to the systems in human-machine interaction via speech.