Lexical and Acoustic Adaptation for Multiple Non-Native English Accents

This work investigates the impact of non-native English accents on the performance of an large vocabulary continuous speech recognition (LVCSR) system. Based on the GlobalPhone corpus [1], a speech corpus was collected consisting of English sentences read by native speakers of Bulgarian, Chinese, German and Indian languages. To accommodate for non-native pronunciations, two directions are followed: Modification of the dictionary to better reflect the non-native pronunciations and adaptation of the acoustic models for native US English with non-native speech. The proposed methods for dictionary modification are data-driven. Therefore no language-specific rules are necessary: The idea is to extract a parallel corpus of phoneme sequences from phonetic transcriptions of native US English and accented English in the George Mason University (GMU) accented database [2]. With this corpus, Statistical Machine Translation models are generated to translate the US English pronunciations in the GlobalPhone dictionary into accented pronunciations which are then used as new pronunciation variants in the GlobalPhone dictionary. With the combination of the lexical and acoustic model approaches, relative improvements of 26.9% for Bulgarian, 33.2% for Chinese, 30.9% for German, and 53.2% for Indian accents are achieved.