Research On Recognition Error Processing In Speech Retrieval

Wih the rapid development of speech recognition technology recently, which has been applied in many fields, speech retrieval technology has emerged. However, due to noise, accent, speed and out-of-vocabulary words, etc., the results of speech recognition will inevitably contain many error messages, which would seriously affect speech retrieval recall and accuracy. In this paper, for the speech retrieval errors arise from speech recognition, we will use error correction, fuzzy retrieval, and other fault-tolerant processing to improve speech retrieval performance. The main work and innovation are described as follows:1. Study of traditional confusion matrix generation algorithmThe main research is on confusion matrix generation algorithm, which calculates confusion probability from1-best and confusion networks. Thus, use the confusion matrix for error correction, the experiments showed that the use of confusion matrix for candidate expansion, can effectively improve the recall rate of correct results. Using approximate matching strategy based on confusion matrix to addressing speech recognition errors. Experiments show that the confusion matrix can effectively recall the query words after faults speech recognition, which improve the recall rate while maintaining recognition accuracy.2. Weighted adjacent word confusion matrix generation algorithmThe traditional confusion matrix generation algorithm can be affected by the alignment errors and error transduction in the recognition results, leading to low accuracy of confusion probability. In this paper, using edit distance and extracting confusion probability from the confusion matrix to determine the recognition of adjacent word, as well as to weight the corresponding word, which ensure the accuracy of confusion probability. Experimental results show that this method is effective to improve the accuracy of the confusion matrix.3. In-depth study selected strategy for confused item on confusion network.For confusion network contains multiple candidates, in the paper will chose the candidates via distinguish right from wrong of the recognition result, set the threshold of posterior probability, and adding time overlap and acoustic score as weights, etc. Comparative experiments are carried out to find the more accuracy confuse items selected strategy.