The Features Extraction And Identification Of Hunan Dialects

The dialect identification technology is applied to judgethe dialect region according to the speaker's pronunciationon the premise that the system know the language belongs to.It's the base of non-standard speech recognition, and importantfor the promotion and application of speech recognition. Therelevant research is not too much currently. The research ofChinese dialect identification is not only conducive toimproving the efficiency of dialect speech recognition system,but also important in the criminal investigation department forpublic security. As a multinational country, it is particularlynecessary to carry out the research of dialects identification.Hunan dialects have selected as research object in thispaper. The features extraction of dialect and the differencebetween dialect characteristics and how to choose appropriateparameter have studied thoroughly. Because the speech signalhas the very strong randomness and the input structure of neuralnetwork is firmly, the dialects identification technologybased on a mixed cascade neural networks of time alignmentnetwork with BP neural network is proposed in this paper, andthe factors which influence identification rate is analyzed.The main work is summarized as follows:1) Extract the dialects acoustics characteristic of HunanChangsha, Zhuzhou, Xiangtan and Hengyang dialects separately,the acoustics characteristic include resonance peak, tonecycle, LPCC coefficient and MFCC coefficient. The differentcharacteristic information of different dialect has analyzedthoroughly in this paper, and the different dialect displays basis which carries on to the dialects identification.2) Took the different characteristic parameter as the inputof the BP network after the time alignment. We discovered thatfor the different dialects and different tone, theidentification rate is not the same when choose differentcharacteristic parameter. The average identification rate isabout 79.2% when took pitch as characteristic parameter, theaverage identification rate is 84.2% when took LPCC coefficientas characteristic parameter, the average identification ratecan reach 86.3% when took MFCC coefficient as characteristicparameter.3) The performance of the system has studied in this paper,and we discussed the influence of alignment number andconcealment level neuron number. The experiment shows that theidentification rate is better when we choose 48 as the alignmentnumber and when the number of concealment level neuron is ten,the performance of the system is better.