IMPROVEMENTS ON MEL-FREQUENCY CEPSTRUM MINIMUM-MEAN-SQUARE-ERROR NOISE SUPPRESSOR FOR ROBUST SPEECH RECOGNITION

Recently we have developed a non-linear feature-domain noise reduction algorithm based on the minimum mean square error (MMSE) criterion on Mel-frequency cepstra (MFCC) for environment-robust speech recognition. Our novel algorithm operates on the power spectral magnitude of the filter-bank's outputs and outperforms the log-MMSE spectral amplitude noise suppressor proposed by Ephraim and Malah in both recognition accuracy and efficiency as demonstrated on the Aurora-3 corpora. This paper serves two purposes. First, we show that the algorithm is effective on large vocabulary tasks with tri-phone acoustic models. Second, we report improvements on the suppression rule of the original MFCC-MMSE noise suppressor by smoothing the gain over the previous frames to prevent the abrupt change of the gain over frames and adjusting gain function based on the noise power so that the suppression is aggressive when the noise level is high and conservative when the noise level is low. We also propose an efficient and effective parameter tuning algorithm named step-adaptive discriminative learning algorithm (SADLA) to adjust the parameters used by the noise tracker and the suppressor. We observed a 46% relative word error (WEB) reduction on an in-house large-vocabulary noisy speech database with a clean trained model, which translates into a 16% relative WER reduction over the original MFCC-MMSE noise suppressor, and 6% relative WER reduction on the Aurora-3 corpora over our original MFCC-MMSE algorithm or 30% relative WER reduction over the CMN baseline.