Research On The Detection Method Of Biases In Mandarin Acquisition Of Ando Tibetan University Students

With the improvement of speech signal processing technology.Tibetan speech recognition technology and speech synthesis technology have made great progress.This brings convenience to the daily life of Tibetan speaker.In recent years,Tibetans have a growing need for learning Mandarin with the economic development and cultural exchange in Tibet.In this influence,it is very important to help Tibetans learn Mandarin.Ando dialect is a Tibetan dialect,which is characterized by no tone,but Mandarin is a tone language.Therefore,Tibetan people who use Ando dialect have difficulties in learning Mandarin.This paper takes the Mandarin pronunciation of the college students who speak Ando dialect as the research object.Research on tone biases and phoneme biases.Combined with experimental phonetics and deep learning,this paper focuses on the detection of tone and phoneme biases.The specific work is as follows:1.Design and establishment of a corpus.The corpus of this paper includes the standard Mandarin corpus and the Mandarin corpus of Ando Tibetan university students.The standard Mandarin corpus is mainly composed of single-tone and double-tone standard Mandarin.The Mandarin corpus of Ando Tibetan university students consists of three parts: 1280 single-tone,2560 double-tone,and 1280 sentences.2.Tone biases detection.The tone biases detection in this article is composed of two parts: Perception experiment and Similarity experiment.In the perception experiment,all corpora in the Mandarin corpus of Ando Tibetan university students were listened to determine whether they were biased,and statistical analysis of results.In the similarity experiment,extract the fundamental frequency of speech and find the similarity of the fundamental frequency curve and average it to get the detection threshold.According to the detection threshold to determine whether there is an error and get the detection result.Combined with perception experiment and similarity experiment to calculate detection accuracy.The results show that when the similarity is used for the error detection of tones,the single-word tone has better detection effect.3.Phoneme biases detection.In this paper,two different acoustic characteristicsare used to detect phoneme biases under different acoustic models.Using the thchs30 corpus as training data,and 1280 sentences of Ando Tibetan university students speaking Mandarin corpus as test data.Under GMM-HMM and DNN-HMM,different acoustic characteristics are used to detect phoneme errors,and the results are evaluated and analyzed.The experimental results show that MFCC has a high detection accuracy under DNN-HMM.