A Study Of Speech Features Extraction And Matching Algorithm Under Noisy Conditions

The speech recognition is one of the most important directions of speech signal processing, its purpose is to remove the communicated barrier between human and machine. Nowadays, speech recognition system can usually get good results under the experimental conditions, but its performance is sharp decline in application. The paper focus on improving the performance of speech recognition systems under noise environment.The paper first introduces the basic overview of speech recognition, describes the significance of the research, and analyzes the speech recognition principle. To improve the recognition rate under noise environment, this paper expounds several common enhancement algorithms, and presents a speech enhancement method based on Hilbert-Huang Transform (HHT), which adaptively adjust threshold according to the different components of basis function. The results of experiments show that it can improve the signal to noise ratio (SNR), speech articulation and intelligibility. And then several common feature parameters for speech recognition are systematically summarized. The advantages and disadvantages of feature based on Linear Prediction Coding (LPC), Perceptual Linear Prediction Cepstrum (PLPC) and Mel Frequency Cepstrum Coefficient (MFCC) are analyzed in detail.In addition, the paper compares the advantages of these methods through experiments. Finally, This paper introduces several pattern-comparison algorithms in detail, including Vector Quantification (VQ), Dynamic Time Warping (DTW) and Hidden Markov Models (HMM). The first two techniques are fit for isolated word and small vocabulary speech recognition systems, achieving good results. However, we usually use HMM in large vocabulary continuous speech recognition system. The search space can be compressed by combining Viterbi algorithm with Beam pruning technique, which means reducing the computational burden. In addition, the paper proposes an optimized method of pruning threshold, and the experiment proves the feasibility of the method.