Abstract
In this paper, we propose the method which improves the speech recognition accuracy on acoustically similar words in Japanese. For example, when there is a different consonant in two acoustically similar words where the sequences of their vowels are same, it is difficult to recognize them correctly under noise circumstances. In the proposed system, after a conventional HMM-based noise robust speech recognition is applied, the acoustic difference in these words is estimated and then noise robust speech recognition is again applied for their precise speech features. As a result, it is shown that the recognition accuracy can be improved under noise environments.
