Диплом (Master), Mississippi State University, 2002, -74 pp.In speech recognition system, it is commonly known that normalization technologies, such as energy normalization and cepstral mean subtraction, they can improve the recognition performance without question. These normalization, n algorithms require the complete utterance to compute the maximum energy and cepstral mean. Thus, it brings a problem into the real-time system. In the Communicator system based on client/server architecture, the audio client sends audio data stream to the speech recognition server. If the speech recognition server has to wait for the complete utterance receiving, it wastes the waiting time and can never be real-time system. We propose some filter based normalization algorithm, which can be implemented as the exponential filter to remove this drawback. We have developed this enhanced algorithm in the Front End of the ISIP prototype system. This enhanced system achieves the same word-error-rate of 3.4% as the Resource Management baseline system. After we make it real-time, it is still comparable to our real-time system with a WER of 5.0%.Introduction Feature Extraction Algorithms Software Design and Implementation Experimentation Impact on the SBE Problem Conclusions and Future Work
Чтобы скачать этот файл зарегистрируйтесь и/или войдите на сайт используя форму сверху.