×

Language recognition based on SVM 1 vs. 1 classification. (Chinese. English summary) Zbl 1299.68127

Summary: Support vector machine (SVM) 1 vs. all classification and the Gaussian back-end classifier are commonly used for language recognition. However, the linear discriminant analysis (LDA) matrix of the Gaussian back-end classifier is often singular so that the performance of traditional 1 vs. 1 classification is worse than that of the 1 vs. all classification. This paper presents an improved 1 vs. 1 classification method that uses re-modeling of the SVM scores. Tests on the the National Institute of Standards and Technology (NIST) 2011 language recognition evaluation (LRE) 30s database indicate that the method gives an equal or even better performance than the traditional method, with linear fusion giving a significantly better performance and with improvements of 7.7% to 15.9% for iVector and 11.2% to 33.9% for the SVM-GSV (Gaussian super vector) language recognition system.

MSC:

68T10 Pattern recognition, speech recognition
68T05 Learning and adaptive systems in artificial intelligence
68T50 Natural language processing
PDFBibTeX XMLCite