A novel approach for degraded speech recognition

Summary form only given. In speech recognition, training and recognition processes can be conducted in different environments. For example, the word pattern may be extracted from clean speech, while the recognition may take place in background noise. Further more, since the vocabulary size is limited, it is possible, in the training phase, to extract the spectral information for all the words to be recognized. The noise reduction filter for each word can thus be constructed from this spectral information