Exploiting High-Level Knowledge Resources for Speech Recognition : Identification, Modeling and Representation of Knowledge Resources for ASR n-best Re-ranking with Applications to Interactive Voice R

ISBN-13: 9783639122121 / Angielski / Miękka / 2009 / 128 str.

Mithun Balakrishna

Opis

This book proposes a novel methodology to improve the§performance of a Large Vocabulary Continuous Speech§Recognizer (LVCSR) by modeling several high-level§knowledge resources into an n-best list re-ranking§mechanism. The book focuses on the identification and§formulation of several novel, additional,§domain-independent knowledge resources into a§re-ranking mechanism. We illustrate the extent of§improvements obtainable by efficiently exploiting§phonetic, lexical, syntactic and semantic knowledge.§We improve WER for specific domains by combining§domain-independent knowledge with automatically§extractable domain-dependent resources. To model§domain-dependent knowledge, we propose a methodology§to automatically generate SLMs for specific dialog§states. The heart of this book not only lies in the§task of selecting and modeling key information§resources but also on combining them efficiently.§Hence, we explore using minimum error rate training§to optimally assign knowledge resource weights by§directly minimizing the WER on a development set.§Finally, we present a novel IVR grammar§creation/tuning application and illustrate the§importance of the re-ranking mechanism in this framework.

O autorze

Mithun Balakrishna received his PhD in Computer Science from The
University of Texas at Dallas. His main fields of research
include automatic speech recognition, spoken language
understanding, and ontology generation from text. He heads the
Spoken Language Technology group and the Ontology/Knowledge-Base
group at Lymba Corporation.