Minimum Phoneme Error based filter bank analysis for speech recognition

In this paper the optimal filter-bank design method based on the Minimum Phone Error (MPE) criteria is investigated. We use Gaussian type filter bank for optimization and various parameters of the filters such as gain, bandwidth and center frequency are trained aiming at maximize the MPE objective function to reduce word error. Preliminary experimental results on a large vocabulary continuous Mandarin speech recognition task given in this paper showed that, compared with both the untrained Gaussian type filters and traditional triangle shaped filter bank, cepstral coefficients derived from the optimized filter bank parameters result in a superior performance for word accuracy. The filters consistent with the MPE criteria are also illustrated.