| The use of speech recognition technology has increased considerably in the last 15 years. However, most of the basic principles and mathematics in signal processing and pattern classification are the same as the ones that were used at the beginning of its market deployment. Furthermore, the accuracy of current state-of-the-art speech recognition systems in special situations like noisy environments and “non-seen speakers in training data” is lower than the recognition accuracy of an average human being.; In this dissertation, I will describe a new approach for handling the signal processing of a voiced speech signal. Rather than using the well-known cepstral coefficients, I will describe a new pitch synchronous approach for analyzing voiced speech signals. This approach consists of using a new pitch tracking method to extract each glottal cycle in the signal and developing a preprocessing stage that normalizes energy and pitch period. Later on, our approach uses component analysis and applies different clustering techniques to get a better probabilistic description of the signal. We describe two new clustering techniques: pitch period clustering and spectral clustering for large training sets.; This new framework is applied to the task of Phoneme Classification in which, given a pitch synchronous glottal cycle, the machine recognizes which phoneme was pronounced. Results show that optimal performance is attained with a very small number of components, requiring low computational power. |
