Speech recognition is a technology which has a rapid development in recent years. It starts to apply to our daily life from the laboratory gradually. However, the recognition rate will decline sharply when the speech recognition engine locates in the complex environment. These noises mainly from the deviation of the channel, the environmental noise, and the mismatch between the contaminated test data and training data etc.Speech recognition is influenced by many factors in the practical application; this article mainly focuses on the impact of noise on speech recognition and how to eliminate this effect. The key to improve speech recognition rate under low SNR environment is endpoint detection, preprocessing stage and feature extraction algorithm.To verify the anti-noise model, I complete speech recognition system based on the CHMM. And I tackled some common problems in CHMM, such as the training set, the choice of model for the initial value and data underflow, etc. From a series of experiment, I found that the performance of the system depends on the phrase of endpoint detection and the signal feature extraction. In order to optimize the anti-noise model, I research various stages of speech recognition, and deepen my understanding of the confrontation noise model. |