earticle

영어

Mandarin Chinese is known as a tonal language with four lexical tones. Tone recognition plays an important role in automatic Chinese speech recognition in that the same syllable with different tones gives quite distinct meanings. The different tone can be characterized by its pitch contour, but the pitch contours are hardly ideal smooth curves. It is because the pitch points calculated by pitch detector normally have some error points. These error pitch points can cause the erroneous classification of Mandarin four-tone recognition. It is necessary to smooth the pitch contour before tone recognition. The classic smooth algorithms can not deal with error fundamental frequencies successively. A new smoothing method proposed in this paper can deal with the error pitch point appropriately. It first checks whether the current point is a correct or error point, then the error type, and finally modifies the error point according to the error type. For different error type, the corresponding smoothing method is also different. To confirm this smoothing method, four “one vs. all” Support Vector Machine classifier are built for Mandarin Tone Recognition. The test results indicate that error rate of Mandarin Chinese four tone recognition can be reduced under the smoothing method.