Based On The The Fisher Quality Chinese Name Voice Recognition Technology

The widely used algorithms now include Vector Quantization(VQ), Gaussian Mixture Model(GMM), Hidden Markov Model(HMM), and so on. These classic algorithms are based on pattern matching, this causes decline of the performance especially in classification of similar Chinese name speech patterns.Support vector machine(SVM) is essentially different with the classic models. It is a kind of discriminative model and has better performance in classification of similar patterns. The classic methods can effectively model dynamic speech patterns, and SVMs have stronger capability in pattern recognition. So, combining the two methods is expected to make use of both advantages. Kernel methods used in SVMs supply us a route to realize this combination. Fisher kernel is the first try to link the probability models and the discriminative classifiers like SVMs, and applied in the detection of biological homology.Fisher score, the key parameter of Fisher kernel, means the feature vectors derived from a probability model. This paper tries to derive Fisher scores from generative models, and uses them in speech recognition. Fisher score can project the variable-length sequence into invariable-length score as feature vectors used to train and test SVMs.This paper investigates three classic models(VQ, GMM, HMM), and develop a speaker verification system on PC. The relationship of them gives us some convenience in the deduction of Fisher scores. We firstly derive Fisher score from simpler GMM, this helps us to further derive it from HMM. For HMM based Fisher score, we take advantage of forward-backward variables used in HMM training. For convenience, the output probability of an HMM is presented by these variables in the form of matrix.Fisher score space is also expanded based on the analyzing of its natural characteristics. New scores are added to the expanded spaces. Different score-spaces have their own physical meanings inspired by Taylor series expansion. Experiments reveal that the new spaces are useful for classification.Experiments of the classic models compare the performances of different...