SGER: Pronunciation Modeling for Conversational Speech Recognition

点击翻译按钮获取中文摘要

Technological advances in the last few years have brought automatic speech recognition where it can now be deployed in limited, real-world applications. However, serious obstacles still remain in satisfactory recognition of spontaneous speech. Casual dialog exhibits a large amount of variability in the pronunciation of words. Humans have little difficulty in transcribing such speech, but automatic transcription suffers badly because of the inability of recognizers to properly consider nonstandard pronunciations. This must be addressed to make systems robust to a conversational speaking style. The investigators propose to examine alternate pronunciations of words in Switchboard, a large corpus of conversational speech, and develop statistical models for predicting the variations based on cues in the acoustic, phonemic and linguistic context of a word. Alternate pronunciations are viewed as perturbations of the canonical ones and, to predict likely variations, the researchers propose to use phonemic environment, lexical stress, syllabic structure, part-of-speech, phrase boundary locations, the realization or deletion of neighboring phones, etc. They shall apply complementary nonparametric and parametric techniques for building predictive models with the goal of improving accuracy in a speech recognition system. This effort addresses a significant obstacle in automatic recognition of spontaneous speech and, in the process, gains fundamental understanding of the causes, mechanisms and extent of pronunciation variability.