A Weighted Overlap Add-based Front-end for Speech Recognition

Speech signal enhancement is frequently referred to as a preprocessing step to speech recognition. However, in practice, this cannot be easily accomplished since the front-end signal processing techniques and/or parameters used in these two frequently differ. We apply a signal processing technique successfully used in speech enhancement to speech recognition and show that it can perform equally well compared to well-known speech recognition front-ends such as MFCC. The technique, oversampled filterbank analysis/synthesis through weighted overlap add (WOLA), has been tested and performed satisfactorily on the TI-46 and Aurora tasks in both clean and noisy conditions and also in subband speech recognition. The results indicate the capability of this technique in reducing the front-end signal processing blocks of enhancement and recognition into a single block.