F0 contour approximation model for a one-stream tonal word recognition system

The performance of a non-tonal speech recognition system degrades when confronted with the task of recognizing tonal words. Several speech recognition applications require tonal word recognition. Therefore, this paper considers how to create a suitable tone model for a tonal syllable recognition system serving application devices based on a one-stream scheme. The fundamental frequency contour (F0 contour) approximation model is proposed here to estimate F0 continuity contours for all of a tonal word. The processes of approximation include voice detection, F0 smoothing, F0 forecasting, and F0 normalization. To model the F0 contours of unvoiced regions belonging to F0 forecasting, a linear regression function is used to create an approximate F0 contour. Experimental results indicate that the proposed model improves the accuracy of tonal word recognition by 8.6% and 12.2%, respectively, compared with conventional random and exponential approaches.