Study Of Speech Recognition Algorithm Based On Modified LP Cepstrum And Neural Network

Speech recognition is a technology with further development in recent years, but because of the complexity of the research, it is a puzzle in the long haul, especially for the speaker-independent speech recognition under noisy environment. Based on a classical speech recognition system, this paper introduced the fundamentals of speech recognition, and discussed several methods in common use of feature extraction, pattern matching, and model training, and carried through improvements on their general algorithms.LP cepstrum is the feature parameters widely used in the technology of speech recognition, but under noisy environment, the recognition rate will decline remarkably. Because Mel frequency scale can reflect sufficiently the nonlinear perceptive characteristic of humans hearing to frequency and amplitude, and frequency analysis and spectrum synthesis characteristics when hearing complex sounds. An improved algorithm for LPCC feature was put forward in this paper, that is to say, LP cepstrum is made nonlinear changes by means of Mel scale according to auditory characteristic, and the LP Mel cepstrum coefficient (LPMCC) is used as feature parameter. The experiment shows that this method is good for robustness and effective on recognition.RBF is a kind of novel, effective forward-feedback neural network.