Research on Chinese Character Confusion Network Algorithm for LVCSR

摘要

In large vocabulary continuous speech recognition, the recognizer outputs using the standard MAP decoding strategy have the minimum sentence error rate, so there is a mismatch between the MAP recognition results and the commonly used performance metric- word error rate. The minimum bayes risk(MBR)decoding strategy can be used to obtain recognition results with minimum WER. One method of MBR decoding is that the word lattice can be transformed into confusion network in order to obtain the hypotheses with minimum WER. According to the characteristic of mandarin, we proposed an Chinese character confusion network generation algorithm based on prevenient works. Firstly, a Chinese word lattice can be produced by using standard mandarin large vocabulary continuous speech recognizer; then the Chinese word lattice is analyzed and handled based on the Chinese language features, and an Chinese character lattice is made; lastly an Chinese character confusion network is produce by implementing alignment in the Chinese character lattice. The experimental results of mandarin large vocabulary continuous speech recognition show that the proposed algorithm yields a lower WER than the MAP recognition and previous two confusion network generation algorithms.