A Study On The Algorithm Of An Isolated Word Recognition System Used In Noisy Environment

In recent years, speech recognition has been a research focus in academic research field, and isolated-word recognition technology occupies an important position in practical applicaton of speech recognition for its advantages of effective computation, small storage and easy realization. Current speech recognition technology, especiallly isolated-word recognition has been relatively mature in theory, and also lots of systems have been built in practical application. However, such systems can only obtain a rather higher accuracy in a lab environment, their performance degrade rapidly in noisy environment, which can not achieve the required recognition rate. Therefore, how to improve the environment adaptability and reduce the impact of noise on the system is the primary consideration for speech recognition system.To solve this problem, the algorithm of an isolated word recognition system used in noisy environment is studied. Many methods are used to reduce noise influence upon recognition rate of the system. This paper focuses on the preprocessing of speech recognition including endpoint detection, feature extraction method and speech enhancement.The main works of this thesis can be summarized as follows:1. Endpoint detection algorithms such as short-time energy, short-time zero-crossing rate and spectral entropy are studied. On the basis of such algorithms, a new voice endpoint detection algorithm based on entropy of correlation coefficients is proposed. To improve the robustness of the algorithm, it adopts enhancement technology of wavelet transform. We simulate and anlyze the endpoint detection algorithms in different noisy environment under different SNR.The simulation result shows that the proposed endpoint detection algorithm has a higher accuracy under low SNR than the algorithm of spectral entropy and basically overcomes the inaccurate in low signal-to-noise ratio environment which exist in the conventional endpoint detection algorithms.2. Searval speech enhancement algorithms such as spectral subtraction, improved spectral subtraction and wiener-filtration are researched.We find that spectral subtraction have large level of musical residual noise, and also improved spectral subtraction has excessive reduction algorithm, although it can suppress the interference of noise. Aiming at the deficiencies of the two methods, an improved algorithm based on wiener-filtration is put forward. The algorithm estimates the priori SNR by introducing cross power spectrum between speech and noise in decision-directed formula.And it also add some processing methods including real time noise updating, voice active detection and some constraints that can enhance speech intelligibility. We simulate the algorithms mentioned above.The result shows that,compared with the other three algorithms, the proposed algorithm can effectively improve the speech quality with the excellent performace in musical residual noise suppression and speech distortion reduction.Eespecially the introduction of speech intelligibility conditions makes intelligibility level of the enhanced voice increase greatly.3. After compared Mel frequency cepstral coefficient (MFCC) with linear prediction cepstral coefficient (LPCC), MFCC with strong robustness is selected as characteristic parameters. Noisy speech signals under different SNR in four representative noisy environmets, such as White noise, Pink noise, Volvo noise and Factory noise are used in the simulations below. First, optimal order of MFCC+ΔMFCC in robustness is selected through experiment. Second, the relative importance of each component in MFCC+ΔMFCC to speech recognition rate in different noisy environment is simulated. The results provided a reasonable basis for reducing the computational time and improving the real-time performance of the system without sacrificing system accuracy. Meanwhile, MFCC+ΔMFCC+LPCC+ΔLPCC and MFCC+ΔMFCC features are simulated in different noisy environment under different SNR, finally MFCC+ΔMFCC are selected by considing speech recognition raio and computation complexity.4. Based on the research of Hidden markov model, an isolated word recognition system with a GUI interface is established. And the definitions of model state number and gauss mixture number are discussed.