Speech Variability Analysis And Application In Isolated Word Recognition

Automatic speech recognition technology has made great progress for the past few years. Along with the spreading of computer, laptop and portable device such as cellular phone, especially the popularization of the Internet, kinds of applications based on speech recognition have sprung up. Specifically, because isolated word recognition has the advantages of high computing efficiency, low storage space and easy implementation, especially in small vocabulary, it is popular in special areas. Although the automatic speech recognition system performs well in the lab environment, its performance degrades a lot in the real application, which is caused by transmission channel, background noise, speaker characteristics and so on. Hence, the speech variability and its effect on speech recognition system have aroused the interests of academia and industry.Surveying the speech variability, isolated word recognition and robust speech recognition, we analyze the variability of transmission channel and speaker emotion state on several corpus with transmission channel and speaker characteristics variability, and try to find the effects of transmission channel and speaker emotion state on speech recognition in feature, model and score domain respectively. In the template based and Hidden Markov Model (HMM) based isolated word recognition frameworks, robust speech recognition technology is explored. In the template based isolated word recognition system, pitch contour remedy algorithm, template selection by cluster method and common vector approach are employed to improve the robustness to speaker emotion variability. In the HMM based isolated word recognition system, cepstral mean subtraction in the feature domain and score normalization in the score domain are adopted to improve the robustness to transmission channel variability. According to the experiment results, these approaches improve the robustness of the isolated word recognition system. Finally, we apply the robust isolated word recognition technology to develop preliminary system applications. This thesis mainly focuses on the following work: 1. Constructing the mandarin isolated word corpus MIWAC.2. Analyzing the effects of the variability of transmission channel and speaker emotion state on speech recognition in feature distribution, model divergence and score distribution respectively on different corpus by kinds of means.3. Exploring the robust isolated word recognition technology to transmission channel and speaker emotion state by using the speech variability analysis result under template based and HMM based frameworks.4. Applying the robust isolated word recognition technology in the information indexing on mobile device and appliance control areas.