Effect of foreign accent on speech recognition in the NATO n-4 corpus

We present results from a series of 151 speech recognition experiments based on the N4 corpus of accented English speech, using a small vocabulary recognition system. These experiments looked at the impact of foreign accent on speech recognition, both within non-native accented English and across different accents, with particular interest in using context free grammar technology to improve callsign identification. Results show that phonetic models built from foreign accented English are not less accurate than native ones at decoding novel data with the same accent. Cross accent recognition experiments show that phonetic models from a given accent group were 1.8 times less accurate in recognizing speech from a different accent. In contrast to other attempts to perform accurate recognition across accents, our approach of training very compact, accent-specific models (less than 3 hours of speech) provided very accurate results without the arduous task of adapting a phonetic dictionary to every accent.