Noise Aware Sub-band Locality Preserving Projection for Robust Speech Recognition

Recovering the nonlinear low dimensional embedding for the speech signals in the clean environment using the manifold learning tech- niques has become of substantial interest recently. However, the issue of manifold learning for feature transformation in domains involving noise corrupted speech can be quite dierent. We tackle this issue by pre- senting a new approach for reducing the noise eect on dierent Mel Frequency Cepstral Coecients (MFCCs) and so Mel sub-bands. We in- troduce our method in the framework of Locality Preserving Projection (LPP) as a manifold learning technique where we construct the manifold on each Mel sub-band by considering the noise eects on it. We name this method as sub-band LPP. we propose to learn one manifold for each MFCC and so Mel sub-band using noisy speech. The experimental re- sults on AURORA-2 database show that the noise aware sub-band LPP improves the noisy speech recognition rate in comparison to conventional LPP for SNR values greater than 0 dB.