Abstract
We describe an efficient method that uses a HMM-based speech synthesis technique as a test pattern generator for evaluating the word recognition rate. The recognition rates of each word and speaker can be evaluated by the synthesized speech by using this method. The parameter generation technique can be formulated as an algorithm that can determine the speech parameter vector sequence O by maximizing P(O¦Q,λ) given the model parameter λ and the state sequence Q, under a dynamic acoustic feature constraint. We conducted recognition experiments to illustrate the validity of the method. Approximately 100 speakers were used to train the speaker dependent models for the speech synthesis used in these experiments, and the synthetic speech was generated as the test patterns for the target speech recognizer. As a result, the recognition rate of the HMM-based synthesized speech shows a good correlation with the recognition rate of the actual speech. Furthermore, we find that our method can predict the speaker recognition rate with approximately 2% error on average. Therefore the evaluation of the speaker recognition rate will be performed automatically by using the proposed method.
