An Improved Tone Model for SVM Based Tone Recognition

Tonal information is important for Chinese Mandarin speech recognition and understanding. Tone information is carried on the shape of fundamental frequencies. When constructing a tone recognition system based on support vector machine, the fundamental frequencies should be normalized by curve fitting method. The fitting coefficients are input data of support vector machine, but it cannot precisely represent the shape of fundamental frequencies. An improved tone model is proposed, which added some other information of fundamental frequencies, such as the maximum and minimum value, the position of maximum and minimum values, the fundamental frequency values of first and last point. The experimental result shows that the proposed model is more accurate with the shape of fundamental frequencies and the accuracy rate of tone recognition is increased by using the improved tone model.