Research On Noise-robust Speech Recognition Based On Feature Extraction

Speech recognition in noisy environments is an important topic and has vast uses in many cases. This dissertation is focused on the problem of noise-robust speech recognition based on feature extraction. While most current speech recognizers give acceptable recognition accuracy for clean speech, their performance degrades drastically when they are applied to real situations, especially in the presence of ambient noise. This has become one of the major obstacles to commercial use of speech recognition techniques. Recent research works concentrate in three categories to deal with this problem. They are noise resistant features and similarity measurement, speech enhancement and speech model compensation for noise. In this thesis, the first two strategies are addressed. The main contributions are listed as follows:1. Proposed an improved method of feature extraction based on spectral subtraction. Because the noise in present frame is replaced by the average of noise in non-speech period in spectral subtraction, errors are caused when the amplitude spectral of clean speech is estimated. The errors will have effects on the feature extraction and the effects can be reduced by the proposed method. Experimental results show that the MFCC feature vectors extracted from noisy speech by the new method are more approximate to the ones extracted from the corresponding clean speech and the recognition rate in noisy environments is raised tremendously.2. Proposed the concept of difference in power spectrum and new noise-robust features based on it. To calculate the difference in power spectrum is equal to remove the additive noise in the power spectrum domain. The power spectrum of the speech signal is allowed to pass through a band-pass filter bank and the differences in the filters' outputs can be the feature vector. Theoretical analysis and experimental results both prove that the recognizer using the proposed features improves its performance greatly in noisy environments.3. Proposed new noise-robust features denoted as OSA-MFCC. The features are extracted from one-sided autocorrelation sequences (OSA), which is the same as...