Improved Large Vocabulary Mandarin Speech Recognition Using Prosodic and Lexical Information in Maximum Entropy Framework

Tone plays an important role in distinguishing ambiguous words in Chinese Mandarin speech recognition. In this paper, we make full use of pitch information. On the one hand, we interpolate F0 contour to make the F0 contour continuous between voiced and unvoiced segments in order to embed F0 into speech recognition system in two streams, which cepstrum and its first and second order derivatives constitute one stream , and F0 and its first and second order derivatives make up the other stream; On the other hand, we use prosodic and lexical features, as well as syllable context information under maximum entropy framework to build explicit tone modeling in rescoring the first-pass outputting lattice. Experimental results show that pitch information and the tonal cues can reduce substitution error greatly and achieve a 3.65% absolute Chinese character error rate (CER) reduction on widely used Mandarin speech recognition tasks-863 test.