Automatic speech recognition: from study to practice

Today, automatic speech recognition (ASR) is widely used for different purposes such as robotics, multimedia, medical and industrial application. Although many researches have been performed in this field in the past decades, there is still a lot of room to work. In order to start working in this area, complete knowledge of ASR systems as well as their weak points and problems is inevitable. Besides that, practical experience improves the theoretical knowledge understanding in a reliable way. Regarding to these facts, in this master thesis, we have first reviewed the principal structure of the standard HMM-based ASR systems from technical point of view. This includes, feature extraction, acoustic modeling, language modeling and decoding. Then, the most significant challenging points in ASR systems is discussed. These challenging points address different internal components characteristics or external agents which affect the ASR systems performance. Furthermore, we have implemented a Spanish language recognizer using HTK toolkit. Finally, two open research lines according to the studies of different sources in the field of ASR has been suggested for future work.