Phonological models in automatic speech recognition
Karen Livescu · 2008
The performance of automatic speech recognition systems varies widely across different contexts. Very good performance can be achieved on single-speaker, large-vocabulary dictation in a clean acoustic environment, as well as on very small vocabulary tasks with much fewer constraints on the speakers and acoustic conditions. In other domains, speech recognition is still far from usable for real-world applications. One domain that is still elusive is that of spontaneous conversational speech. This type of speech poses a number of challenges, such as the presence of disfluencies, a mix of speech and non-speech sounds such as laughter, and extreme variation in pronunciation. In this talk, I will focus on the challenge of pronunciation variation. A number of analyses suggest that this variability is responsible for a large part of the drop in recognition performance between read (dictated) speech and conversational speech.