The Research On Segmentation Acoustic Model Based On MPE Tibetan Lhasa Dialect

Acoustic model is one of the most important elements in speech recognition.Its accuracy has a direct impact on the recognition effect of speech recognition system.Building a more accurate acoustic model has been the major focus of this discipline.In light of this,we will look into the estimation processes of three tone parameters based on the minimum phoneme error criterion in order to improve the accuracy of acoustic model.The key of continuous speech recognition in large vocabulary is to establish and train an acoustic model which can describe the acoustic characteristics accurately.Thus the choice of training criterion has a significant impact on the recognition rate.There are many types of training methods for acoustic models.The traditional training methods can only train the inner part of the model,but the models cannot distinguish from each other.Given this limitation,discriminative training is usually applied to solve this problem.Compared with the traditional training methods,discriminative training algorithm includes the information on the boundary of the model.Therefore the acoustic model with better recognition performance can be obtained.In this paper,the discriminative training of acoustic model is investigated using the continuous speech recognition system in large vocabulary based on Tibetan Lhasa dialect.The detailed research contents will be further explained in below.In this paper,the traditional maximum likelihood estimation training algorithm based on generative criterion and minimum training algorithm based on discriminative training criterion(Minimum,Phone,Error,MPE)are studied.The experiment platform of each training algorithm is built by using HTK tool,and the Tibetan Lhasa dialect acoustic model based on these two methods is established.A total of five experiments,experiment 1 three triphone models have better recognition effect by selecting the experimental modeling unit;Experiment 2 by setting the Gauss mixture for different number of Experiment 3 was to verify;through the penalty factor set,get the need to find a critical value to improve the recognition effect;Experiment 4 the Phone Lattice set the size to set according to the actual situation;Experiment 5 is whether to join the I-smoothing function,have joined this smoothing function,better recognition effect.The experimental results show that the minimum phoneme error training algorithm improves the phoneme recognition rate compared with the traditional generative acoustic model training method.Compared with the maximum likelihood estimation criterion,the correct recognition rate of single syllable is improved by 7.15%,and the correct recognition rate of three tones is increased by 7.78%.