Exploring low-dimensional structures of modulation spectra for robust speech recognition

Developments of noise robustness techniques are vital to the success of automatic speech recognition (ASR) systems in face of varying sources of environmental interference. Recent studies have shown that exploring low-dimensional structures of speech features can yield good robustness. Along this vein, research on low-rank representation (LRR), which considers the intrinsic structures of speech features lying on some low dimensional subspaces, has gained considerable interest from the ASR community. When speech features are contaminated with various types of environmental noise, its corresponding modulation spectra can be regarded as superpositions of unstructured sparse noise over the inherent linguistic information. As such, we in this paper endeavor to explore the low dimensional structures of modulation spectra, in the hope to obtain more noise-robust speech features. The main contribution is that we propose a novel use of the LRR-based method to discover the subspace structures of modulation spectra, thereby alleviating the negative effects of noise interference. Furthermore, we also extensively compare our approach with several well-practiced feature-based normalization methods. All experiments were conducted and verified on the Aurora-4 database and task. The empirical results show that the proposed LRR-based method can provide significant word error reductions for a typical DNN-HMM hybrid ASR system.