Research on confusion network algorithm for Mandarin large vocabulary continuous speech recognition

Word error rate (WER) is usually used as the performance evaluation metric in speech recognition. However, decoding in most state-of-art speech recognizer is based on the maximum a posterior probability (MAP) decision rule, and recognition results have the minimum sentence error rate. For the purpose of keeping the identity between decoding criterion and performance measurement, we principally discuss the minimum WER decision rule of speech recognition decoding. This means that Levenshtein distance is used as loss function of Bayes decision theory in order for decoder to seek the speech recognition results with minimum Bayes risk. To compute the Levenshtein distance, one method is to transform word lattice output from speech recognizer into confusion network for consensus decoding. According to the characteristic of Chinese lexicon, we proposed a Chinese character confusion network algorithm based minimum WER decision rule used in mandarin large vocabulary continuous speech recognition. The experimental results based on 2005 HTRDP (863) Evaluation Corpus show that the proposed algorithm yields a lower WER than the MAP recognition and word confusion network decoding.