Bidirectional OM-LSA speech estimator for noise robust speech recognition

A new speech enhancement method using bidirectional speech estimator is introduced. A widely-known speech enhancement method using the optimally-modified log spectral amplitude (OM-LSA) speech estimator is re-modified under the assumption that the frame-synchronous estimation is not essential in some of the speech recognition applications. The new method utilizes two separate flows of the speech gain estimation, one is along the forward direction of time and the other along the backward direction. A simple look-ahead estimation mechanism is also implemented in each flow. By taking the average of these two gains, the speech estimation becomes more robust under various noise conditions. Evaluation experiments using the artificial and real noisy speech data confirm that the speech recognition accuracy can be greatly improved by the proposed method.