Symbolic Speaker Adaptation for Pronunciation Modeling

This paper presents a method of modeling a speaker’s pronunciation of a given language as a blend of ”standard” speech and other non-standard speech varieties (regional dialects and foreign accented pronunciation styles) by way of speaker-dependent modification of a lexicon. In this system, a lexicon of Standard American English (SAE) forms, the ”canonical” lexicon, is filtered and transformed via a group of speech variety (SV) dependent rule sets into a speaker specific set of pronunciation variants (and associated probabilities) for use during recognition. The relative importance of these rule sets depends on the speaker’s pronunciation characteristics and is represented by a Speech Variety Profile (SVP) associated with each speaker. A speaker’s individual SVP is acquired through feedback from an adaptation process. Convergence to a speaker’s SVP represents adaptation of the lexicon (symbolic adaptation) to those SVspecific forms that speaker is likely to utter.