全局信息融合的汉语方言自动辨识

提出身份认证矢量（Identity vector，I-vector）结合韵律信息的汉语方言辨识方法。全差异空间替代本征音与本征信道空间，将高维超矢量映射为低维I-vector表示，并进行信道补偿与特征降维处理。汉语是有调语言，各方言在其韵律结构上具有明显差异，I-vector特征融合全局韵律信息，可有效增补各方言鉴别性。利用融合信息对闽、粤、吴等五种方言以及普通话进行辨识实验，等错率（Equal Error Rate，EER）达到8.01%，比高斯混合模型-通用背景模型（Gaussian Mixture Model-Universal Background Model，GMM-UBM）降低56.2%，表明融合全局韵律信息的I-vector方法可有效提高汉语方言辨识正确率。

Automatic identification of Chinese dialects based on global infor-mation fusion

A new method of Chinese dialects identification based on Identity vector（I-vector） combined with prosodic information is proposed. The high-dimensional super-vector is mapped to a low-dimensional I-vector representation by Total Variability（TV） model. Channel compensation and feature dimension reduction are also performed. Chinese is a typical language with a tone and Chinese dialects have obvious differences among rhythm, stress and other rhythmic structure. The serial fusion of I-vectors with global prosodic information can improve the distinguishability of Chinese dialects effectively. The Equal Error Rate（EER） using fusion strategy of five Chinese dialects and Mandarin is 8.01%, which is 56.2% lower than the Gaussian Mixture Model-Universal Background Model（GMM-UBM） method. The experimental results show that the I-vector method fusing global prosodic information can improve the Chinese dialects identification accuracy effectively.