A speech enhancement preprocessor for robust speech recognition

To improve the robustness of automatic speech recognition (ASR) system in adverse environments, speech enhancement preprocessor has been widely used recently to reduce the impact of noise. In this paper, an improved noise estimation approach is proposed in the enhancement preprocessor to keep enhancement performance with low distortion and complexity. First, the noisy speech is transformed into Bark domain and the minima values are tracked. Then, these values are applied to control the smoothing factor of recursive averaging in noise estimator. Based on perceptual model, the proposed approach makes full use of correlations among frequency bins in the same Bark band to track noise better. Furthermore, this approach has low complexity and is suitable for consecutive recognition process. Simulation results show that the proposed preprocessor can effectively improve recognition accuracies at very low SNRs.