Study on the integration of speech and language processing in recognition of Chinese continuous speech

This paper presents a method of Chinese continuous speech recognition, which synthesizes speech and language processing with frame-synchronous parsing algorithm. The evolved language processing employs the context free grammar and top down sentence analyzer. The evolved speech processing uses the One Pass Viterbi algorithm based on finite state automaton. In the evaluation experiments, 60 phonemic HMMs and 8 tone HMMs were used. By using the proposed algorithm, we obtained the average sentence recognition rate of 94.4% for 1070 utterances of ten speakers, and an improvement of 8% is obtained in the same task of perplexity 27.3, compared to conventional hierarchical system based on word spotting and lattice parsing algorithm.