Candidate generation for interactive Chinese speech recognition

Abstract

Despite progress in automatic speech recognition, user involvement and assistance are still needed to achieve performance similar to human in real-world tasks. Though many interactive speech recognition systems using candidate selection are developed for English and Japanese, similar systems for Chinese face particular difficulties due to characteristics of the Chinese language, especially at the candidate generation stage. In this paper, a method is proposed for candidate generation in interactive Chinese speech recognition, which lists all competitive candidates in forms of Chinese characters in stead of words. The main idea is to align time-overlapped links in Chinese word lattice into clusters based on their phonetic similarity, and then obtain candidates by splitting words attached to links into Chinese characters. Experimental results demonstrated that the candidates generated by the proposed method can lead to high performance in interactive Chinese speech recognition.