| Language Recognition is used widely. It has abundant content, utility and interest. This makes it be a hot research domination and be used many fields such as language inquiring in office or business system, language control in industry, language dialing in telephone and telecommunication system medical and sanitation system and so on. The mature techniques include HMM (Hidden Markov Model), V.Q (Vector measure),DTW (Dynamic Time Wrap).An Isolated word Language Recognition system with a small list for special person is built in this paper, which includes signal processing including orientation of language beginning and end, feature extraction, data condense and recognition.The orientation of language beginning and end with white noise is studied in this paper. A method of a series Meier filters is experimented, which is the usage of the nonlinear filter character of human ear and the essential of the likeness between a common noise and a white noise. In this paper, it is compared with the method of short-time amplitude, improved the method of short-time amplitude, the method of microwave transfer. From the experiments, with the original short-time amplitude methods, a proper error rate(3.83%) could be got with proper threshold without additional noise ,but it is not robust, when the energy ratio between signal and noise is smaller than 10dB,the methods is invalidation. the improved short-time amplitude which uses a changeable statistic threshold in the place of a fixed threshold is more robust than the original one, but its error rate is still high, it respective is 21.53% and 29.02% with the energy ratio between signal and noise 10dB and 5dB and the methods is complicate. the methods with microwave transfer is good at looking for sonant's beginning and end, it needs a more accurate methods to look for surd. The orientation of beginning and end required with the methods with a series MEIER filters is comparatively accurate than the others mentioned in this paper, it respective is 11.96% and 18.17% with the energy ratio between signal and noise 10dB and 5dB ..The result illuminates that the method of a series Meier filters is more efficient and robust than others. As far as the feature extraction, the traditional MEIER Cepstum is improved by adding a difference filter before handling and harmonization after handling to decrease the influence of noise. The experiment illustrates that the improved method can make sure the correct rate of word recognition under the strong noise background. In the comparative experiment of the energy ratio between signal and noise between The MEIER Cepstum (MFCC) feature and the improved MEIER Cepstrum (DPSCC) feature extracted from Chinese word"YOUZHUANG"with additional white noise, whose energy ratio between signal and noise is 5dB is respective 4dB and 7dB. In the comparative experiment of recognition correct rate through DHMM model with the two different features using the original language and adding the white noise whose energy ratio between signal and noise are respective 20dB, 10dB, 5dB. For the MFCC, its correct rate is respective 96.22%, 95.11%, 89.00% and 87.33%,and for the DPSCC, its correct rate is respective97.33%, 97.33%, 91.78%and 90.11%.The methods of PCA (primary component analysis) and LBG VQ (vector quantification) are adopted to condense the data. The Euclid distance between vector and the code word is used as the distortion measure, the ratio between the relative decrease of two continual total distortions and the former total distortion (smaller than 0.0001)[35] or the biggest circle number (1000) is used as the end condition of LBG circle. The number of code is 1028.The initial code is acquired through separating method. Through the PCA, the dimension of feature is decreased to 16 from 31; The variance devotion of the 16 feature is 0.99.A DHMM (Discrete Hidden Markov Model) is set. The number of state, the training end condition, the initialization of parameter and the overflow of data are studied or illuminated in the model. a five-status left-right model allowed one status jump model is adopted. During the training of DHMM with the method of Baum-Welch, using the output probability logarithm, which is calculated with the method of Viterb,prevents its overflow and using the sum of former front probability as the ratio gene prevents the front or behind probability's overflow. the initial value of B( the matrix of observation vector probability) is calculated by the means of segmental K average and the initial value of matrix A( the matrix of status transfer probability) is supposed to averaging distribution. Using the model builds respectively model for every word. The VQ_HMM method is also be studied as the improved one of DHMM. The experiment illustrates the latter is simpler without compute the observation possibility matrix and can make sure higher correct rate of recognition. The correct rate of recognition with training material and test material is respective 98.17% and 97.33% with model of DHMM,as the comparative rate, the correct rate of recognition with training material and test material is respective 98.44% and 97.36% with model of VQ-HMM. |
