Feature extraction is one of the active research areas of speech recognition. Local features have received great attention among speech recognition researchers, since local features have advantages over the traditional speech features for reducing the effect of noise on the recognition performance. Here, our main focus is on extracting local features for speech recognition. Two new speech features are proposed for speech recognition. The first one is the lapped subband (LAP) features, extracted using the discrete orthogonal lapped transform. The second one is the Mel-Frequency Discrete Wavelet Coefficients (MFDWCs), extracted using the Discrete Wavelet Transform (DWT). The proposed features take advantage of being local in frequency domain which makes them robust to the noise.; Performance of the LAP features was evaluated on a phoneme recognition task and compared with the performance of subband (SUB) features and Mel-Frequency Cepstral Coefficients (MFCCs) features. Experimental results have shown that the proposed LAP features outperform SUB features and MFCCs features under white noise, band-limited white noise and no noise conditions. We evaluated the performance of MFDWCs for clean speech and noisy speech and compared the performance of MFDWCs with MFCCs, SUB features and multi-resolution (MULT) features. Experimental results on a phoneme recognition task showed that a MFDWC-based recognizer gave better results than recognizers based on MFCCs, SUB, and MULT features for the white Gaussian noise, band-limited white Gaussian noise and clean speech cases.; The Parallel Model Combination (PMC) technique is used along with MFDWCs features to take advantage of both noise compensation and local features (MFDWCs) to decrease the effect of noise on recognition performance. In addition a practical weighting technique based on the noise level of each coefficient is introduced. The experimental results show significant performance improvements for MFDWCs versus MFCCs after compensating the HMMs using the PMC technique. Weighting the partial score of each coefficient based on the noise level further improves the performance.; This work also investigated application of the DWT to the time-trajectory of the static coefficients to extract dynamic features.
展开▼