Phonetic Event-based Whole-Word Modeling Approaches for Speech Recognition

Speech is composed of basic speech sounds called phonemes, and these subword units are the foundation of most speech recognition systems. While detailed acoustic models of phones (and phone sequences) are common, most recognizers model words themselves as a simple concatenation of phonemes and do not closely model the temporal relationships between phonemes within words. Human speech production is constrained by the movement of speech articulators, and there is abundant evidence to indicate that human speech recognition is inextricably linked to the temporal patterns of speech sounds. Structures such as the hidden Markov model (HMM) have proved extremely useful and effective because they offer a convenient framework for combining acoustic modeling of phones with powerful probabilistic language models. However, this convenience masks deficiencies in temporal modeling. Additionally, robust recognition requires complex automatic speech recognition (ASR) systems and entails non-trivial computational costs. As an alternative, we extend previous work on the point process model (PPM) for keyword spotting, an approach to speech recognition expressly based on whole-word modeling of the temporal relations of phonetic events. In our research, we have investigated and advanced a number of major components of this system. First, we have considered alternate methods of determining phonetic events from phone posteriorgrams. We have introduced several parametric approaches to modeling intra-word phonetic timing distributions which allow us to cope with data sparsity issues. We have substantially improved algorithms used to compute keyword detections, capitalizing on the sparse nature of the phonetic input which permits the system to be scaled to large data sets. We have considered enhanced CART-based modeling of phonetic timing distributions based on related text-to-speech synthesis work. Lastly, we have developed a point process based spoken term detection system and applied it to the conversational telephone speech task of the 2006 NIST Spoken Term Detection evaluation. We demonstrate the PPM system to be competitive with state-of-the-art phonetic search systems while requiring significantly fewer computational resources