Environment Compensation For Speech Recognition

Although speech recognition technique develops fast after many years' research, there are many problems to be solved urgently. In particular, the adverse noise environment greatly restricts this technique for application. We can categorize many state-of-the-art environment compensation approaches into two classes: the front-end and the back-end (acoustic model) processing approaches. Currently, the environment compensation approaches based on statistical model are paid a great attention and have been successfully applied in the front-end and back-end. Experimental results show that they can greatly improve the system performance compared to traditional approaches. In this thesis, we will make a deep research based on our speaker-independent large vocabulary continuous speech recognition (SI-LVCSR) system. The SI-LVCSR system is composed of three open source toolkits. Acoustic model is trained by Hidden Markov Toolkit that is developed by Cambridge University. Language model is trained by CMU-Cam_Toolkit that is developed by Carnegie Mellon University and Cambridge University. The recognizer is Julius that is developed by Kyoto University and IPA.The research and innovations are described in details as follows:1. Based on a deep research on the speech corruption process, we respectively establish environment models for time domain, linear spectral domain, log spectral domain and cepstral domain.2. We establish a feature compensation approach based on batch EM noise estimation algorithm, and respectively formulate two different...