Abstract
This thesis presents a research on recognizing the retroflex and non-retroflex speech sounds for Mandarin Chinese. The objective of this research is to identify whether the initial of a syllable obtained from forced alignment is retroflex or not.Hidden Markov model-based (HMM) retroflex and non-retroflex models are used in this research. Besides the conventional speech features like Mel-scale frequency cepstral coefficients (MFCC) and log-Energy, the energy ratio taken from the initial segment is also used. Different energy ratios are obtained by adjusting the proportions of the low-frequency part to the high-frequency part. Different models are also constructed based on different modeling approaches. The final of a syllable is also used for training. Viterbi algorithm is used for recognition, and the recognition network is adjusted according to different modeling approaches.The experimental results show that the best performance is reached when using the energy ratios of the front 25% part to the tailing 75% part of the spectrum in conjunction with the MFCC and log energy features. A 98.24% recognition rate on retroflex and non-retroflex speech sounds for Mandarin Chinese is obtained.