【摘要】 目的以方言语音学为视角,通过计算机分析和选择模型训练数据,探究提高口音自动识别准确率及优化模型训练数据的方法。方法采用音系分析、语音信号处理、数学模型实验和统计的方法,对包含约81400段电子语音的37个方言点(每个方言点约2 200段),逐一进行音系提炼、挑选音系例字、电子语音预处理和提取梅尔倒谱系数(MFCC),构建高斯混合模型(GMM)进行口音识别分析。结果从方言语音中提取的音系例字作为训练集(约260字)的识别模型可以较好地完成口音识别任务,对比任选300个例字作为训练集的识别模型,无论是识别准确度还是待测语音数量的要求都有显著优势。结论基于语言学音系例字的口音自动识别方法,已经完成37个方言点高斯混合模型的建立,可以用于口音识别中辅助辨别分析。更多还原

【Abstract】 Objective From perspective of dialect phonetics, through computer analysis and selection of model training data,this study explores the methods to improve the accuracy of automatic accent recognition and optimize the model training data.Methods By using methods of phonological analysis, signal processing, mathematical model and statistics, 37 dialect points(each dialect point is about 2 200) containing about 81 400 segments of electronic recording are selected phonological words,electronic speech preprocessing, extracted Mel-scale frequency cepstral coefficients, and Gaussian mixture models are constructed for accent recognition and analysis. Results The results show that the recognition model based on the training set of260 words extracted from dialect speech can complete the task of accent recognition. Compared with the recognition model based on the training set of 300 words, it has significant advantages in recognition accuracy and the number of speech to be tested. Conclusion The automatic accent recognition method based on linguistic phonological words has completed the establishment of Gaussian mixture model of 37 dialect points, which can be used for auxiliary discrimination analysis in accent recognition.更多还原