This study proposes a novel missing-feature reconstruction method to improve speech recognition in background noise environments. In order to improve the existing missing-feature reconstruction method which utilizes only frequency correlation, a temporal spectral feature analysis is employed to leverage temporal correlation across neighboring frames. The final estimates for missing-feature reconstruction are obtained by a selective combination of the frequency correlation based method and the proposed temporal correlation based method. Performance of the proposed method is evaluated using the Aurora 2.0 framework with car noise and speech babble conditions. Experimental results demonstrate that the proposed method is more effective at increasing speech recognition performance in adverse conditions. By employing the proposed temporal-frequency based reconstruction method with SNR-based mask estimation, +21.31% and +20.73% average relative improvements in WER are obtained for car and speech babble conditions, compared to the original frequency correlation based method.
展开▼
