A novel classifier modification approach to missing data problem for noisy speech recognition

Missing data theory has recently been used as a solution to noise robustness issue in Automatic Speech Recognition (ASR). Missing components of spectrogram can either be reconstructed, as carried out in Spectral Imputation, or simply ignored, as done in classifier modification. Most of the research has been focused on imputation because of the problems associated with classifier modification approaches. In order to address these issues, we propose to transfer Bounded Marginalization (BM), the classic classifier modification approach, to cepstral domain, employing a proposed uncertainty transfer function. We have named the proposed technique as Bounded Cepstral Marginalization. Our proposed approach has shown better recognition accuracy than BM, even by employing smaller feature vectors. Also, it shows better robustness while employing an inaccurate estimated mask.