Normalization on Temporal Modulation Transfer Function for Robust Speech Recognition

In this paper, we proposed a robust speech feature extraction algorithm for automatic speech recognition which reduced the noise effect in the temporal modulation domain. The proposed algorithm has two steps to deal with the time series of cepstral coefficients. The first step adopted a modulation contrast normalization to normalize the temporal modulation contrast of both clean and noisy speech to be in the same range. The second step adopted an edge-preserved smoothing to attenuate the low modulation components while preserving the high modulation components (edges). We tested our algorithms on speech recognition experiments in both additive noise condition (AURORA-2J data corpus) and reverberant noise condition (convolution of clean speech utterances from AURORA-2J with a smart room impulse response signal). For comparison, the ETSI advanced front-end algorithm (AFE) is used. Our results showed that the algorithm got: (1) for additive noise, 57.26% relative word error reduction (RWER) rate for clean conditional training (59.37% for AFE), and 33.52% RWER rate for multi-conditional training (35.77% for AFE), and (2) for reverberant noise, 51.28% RWER rate (10.17% for AFE).