Research On Recognition Methods Of Hunan Dialects Based On BP_Adaboost And HMM

Nowadays, Chinese is spoken by the largest number of people and distributed second widely in the world.The world is on the rise of an "Chinese fever" with China’s increasing global influence.China,as the Chinese population than any other country.With the growing domestic economic development and the pace of urbanization speeding up,People communicate with each other become more frequent.Some like public information security, linguistic engineering,intelligent retrieval,public facilitiesbecoming perfect and demanding for dialect recognition technology extremely. The potential and opportunities of the Chinese dialect identification technology will be very broad in the future.This paper focuses on the non-specific isolated word recognition and identification of Hunan regional dialects.In order to reflect the dynamic properties of dialects and the characteristics of vocal tract, Use the LPCC,MFCC and their first order differential factors combined into the48-dimensional mixed characteristic parameters in order to represent the dialects signal.Improve the recognition rate of the Chinese dialect identification system and anti-noise performance.In this paper, on the basis of the elementary theory and algorithm of the Adaboost,BPNN and HMM,dialects of recognition and identification methods have been proposed.A Chinese dialect identification method is a combination of BPNN and Adaboost machine learning algorithm.BP neural networks used to be a weak classifier dialect for initial identification. The BPNN combined constitute a classification accuracy of higher intensity classification by Adaboost iterative algorithm. This dialect identification method uses the adaptive weights of the adaboost to enhance BPNN classification capabilities. Through the establishment in the different signal-to-noise ratio, the different characteristic parameter, in the different identification model situation recognizes rate the contrast experiment.Results show that mixed has better features than other characteristic parameter of characterization and noise immunity.Based on the BP Adaboost identification method also has a higher literacy rate.Another Chinese dialect identification method is a combination of hidden Markov model and BP_Adaboost model.Known by the first method to the words of a dialect isolated geographical information.Primary by using Baum-Welch,Viterbi algorithm for training and recognition.The second identification, and comprehensive final results by bp_adaboost. Identifies the contrast parameter library derived its specific meaning. This hybrid model is fully absorbed the time series modeling capabilities of the HMM model and strong classification ability of the integrated neural network. The experimental results show,these two methods have a stronger noise robustness and a high recognition rate than BP neural network or HMM.