Subword lexical modelling for speech recognition

In this work, we introduce and develop a novel framework, scANGIE, for modelling subword lexical phenomena in speech recognition. Our framework provides a flexible and powerful mechanism for capturing morphology, syllabification, phonology, and other subword effects in a hierarchical manner which maximizes sharing of subword structures. A scNGIE models the subword structure within a context-free grammar and an accompanying probability model. We believe that our framework has several advantages: The sharing mechanism allows training data to be pooled amongst instances of the same word substructure even when they occur across different words in the lexicon. Further, knowledge of this substructure can be extended to filler models in a word-spotter, new words added incrementally to a recognizer's vocabulary, and potentially in support of new word detection. The context-free foundation allows for ease of research and experimentation with varying subword representations, and also facilitates integration with a natural language understanding system. Finally, the availability of subword structural information in a recognition system enables exploration of prosodic models which use this information. In this thesis, we demonstrate scANGIE's feasibility and efficacy in a variety of applications. Using scATIS corpus data, we show that scANGIE results in performance improvements on phonetic recognition, reducing error rate from 39.8% to 36.1% as compared to a phone bigram baseline. We show its competitiveness in the task of word-spotting, where we also report on a comparative study of different subword lexical models for the filler space. The FOM results ranged from 85.3 for a phone bigram to 89.3 for a system using the full scANGIE parse tree and a lexicon of 1200 words. We also discuss an implementation of a competitive continuous speech recognition system based on scANGIE, which achieves a recognition error rate of 18.8% on our test set as compared to a baseline error rate of 18.9%, both using a word bigram. Finally, we explore the integration of scANGIE with a natural language understanding system, resulting in a fully coupled system, based on context-free frameworks for both phonological and linguistic modelling. The integrated system achieves a recognition error rate of 14.8% on the same test, an improvement of 21.6%. We will also discuss two pilot studies, one on handling dynamic vocabulary updates within a continuous speech recognizer and the second on hierarchical duration modelling within a word-spotter. Both studies showed promising results. (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge, MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.)