Research On Speech Recognition Algorithm In A Noisy Environment

With the development of science and technology, voice recognition algorithms and the corresponding recognition systems are already quite mature, including Dynamic Time Warping algorithm (DTW), Vector Quantization algorithm(VQ), Hidden Markov Model (HMM), Artificial Neural Network(ANN) and other algorithms. These algorithms make speech recognition highly improved in both the recognition rate and the recognition efficiency. Currently, in the laboratory environment, speech recognition systems for clean speech can reach more than 95% of the recognition rate, with high recognition accuracy.However, in certain circumstances, with various sources of noise, such as noise in the environment, the speech recognition rate of the system will be greatly affected. Since the training samples with the sample library feature mismatch recognition performance of the system,so that these systems have a lot of the sharp decline. Therefore, if we want the speech recognition system to be used effectively, we have to design the identification of the anti-noise which is really needed to be solved.Firstly, the development of speech recognition in recent decades are introduced, including the voice signal sampled separation de-noising, pre-emphasis, sub-frame window, end-point detection, etc. The procedure is described in detail on the end-point detection by two methods described for analysis. After that, the on-board voice recognition system is introduced, with the presentation of the concept of battle command vehicle car speech recognition systems. By using the bus environment to give a simulation to verify the performance of de-noising algorithm.Then from the research of de-noising technology, we study signal separation method in the case of the presence of noise speech We focus on the blind signal separation based on tabu search algorithm of noise technology, and design an experiment in the waveform signal separation to examin the performance of FastICA algorithm. Then we choose FastICA as our demising experimentNext article describes the extraction of characteristic parameters of the speech signal We contrast LPC coefficients, LPCC coefficient and MFCC coefficient, then we select one coefficient in the final experiment. After performing feature extraction, we introduces techniques for vector quantization, focusing on algorithm design LBG codebook.Besides, HMM-based speech recognition technology has been studied, including the introduction of the HMM and three problems that exist in speech recognition systems. By solving these problems, we implement a voice recognition system based on HMM. Then Multi-stage HMM algorithm is proposed to design a voice recognition system to improve the recognition rate of existing systems.At last we select four small library of voice recognition system performance tests in four different environments, including the selection of characteristic parameters, the code number of the capacity and determination of HMM states, as well as the effects of noise in different environments affected the size of the environment and so on. After that, by comparing with the dynamic time warping algorithm and enhanced voice based on wavelet transform algorithm in different noise environments, we prove the improvement of the proposed algorithm.