SPEAKER ADAPTATION USING COMBINED TRANSFORMATION AND BAYESIAN METHODS

Adapting the parameters of a statistical speaker-independent continuou s-speech recognizer to the speaker and the channel can significantly i mprove the recognition performance and robustness of the system. In co ntinuous mixture-density hidden Markov models the number of component densities is typically very large, and it may not be feasible to acqui re a sufficient amount of adaptation data for robust maximum-likelihoo d estimates. To Solve this problem, we have recently proposed a constr ained estimation technique for Gaussian mixture densities. To improve the behavior of our adaptation scheme for large amounts of adaptation data, we combine it here with Bayesian techniques. We evaluate our alg orithms on the large-vocabulary Wall Street Journal corpus for nonnati ve speakers of American English. The recognition error rate is approxi mately halved with only a small amount of adaptation data, and it appr oaches the speaker-independent accuracy achieved for native speakers.