Application Research On Speech Recognition Algorithms For Interactive Game Engine

Speech recognition system for interactive game engine is used to make the computer understand some speech orders and drive the model react correspondingly. The function involves two kinds of technology: speech recognition and 3D-model, and speech recognition is the most important technology, which is crucial for the result of interaction between man and computer. So it is necessary to research the speech recognition system.Speech recognition system includes several parts: endpoint detection, feature extraction, and pattern matching. Endpoint detection is used to gauge the starting frame and ending frame of speech signal and differentiate from noise signal frame. Feature extraction is used to extract some parameters which can be used to stand for the character of speech signal. Pattern matching is used to compute the similarity between input signal and template. In this paper a new endpoint detection algorithm based on pitch frequency is suggested. The results of experiments show that the algorithm is more real-time and more robust. Another endpoint detection algorithm is improved based on the theory of C₀ complexity. The new detection method improves the traditional C₀ complexity and combines with enhanced short-time energy. Simulation results indicate that the method has a strong robust performance and can detect the endpoint of speech even in low SNR situation and can be applied to real-time speech recognition system.A new kind of speech signal feature: second-order weighted Mel Frequency Cepstrum Coefficients is also suggested in the paper by introducing the method of weighted filter bank analysis. It also studies the weighted factor in the process of obtaining the Inverted Mel Frequency Cepstrum Coefficients and presents the optimal weighted factor. The result of the experiments shows that this new parameter is more robust, and can be applied to generic speech recognition system. Besides two kinds of pattern matching methods, DTW, HMM are analyzed deeply. Through analysis and experiment, the appropriate algorithm is choosed and applied to the recognition system successfully.