Phonetic Variation Analysis Via Multi-Factor Sparse Plus Low Rank Language Model

Phonetic transcriptions contain rich information about language. First, the sequential patterns in phonetic transcripts reveal information about the language’s phonotactics. When combined with lexical information, this can help to grow or correct pronunciation dictionaries and to improve grapheme-to-phoneme prediction. Second, the places where pronunciations deviate from the norm can be equally informative; for example, by providing cues for speaker traits such as accent, dialect or sociolect. Interesting in itself, detecting speaker characteristics can also be used to improve speech recognition system performance (Biadsy, 2011). In this extended abstract we describe on-going work to automatically analyze both the regularities and the exceptions (deviations) in phonetic sequences. We use the Multi-Factor Sparse Plus Low Rank Language Model (Hutchinson et al., 2013), which was shown to effectively model regularities and exceptions in word sequences (e.g. by identifying lexical deviations characteristic of topic or speaker role). Preliminary results modeling commonalities and variation between dialects of American English are promising and suggest several extensions to this work.