It is well known that speaker variability caused by accent is one fact
or that degrades performance of speech recognition algorithms. If know
ledge of speaker accent can be estimated accurately, then a modified s
et of recognition models which addresses speaker accent could be emplo
yed to increase recognition accuracy. In this study, the problem of la
nguage accent classification in American English is considered. A data
base of foreign language accent is established that consists of words
and phrases that are known to be sensitive to accent. Next, isolated w
ord and phoneme based accent classification algorithms are developed.
The feature set under consideration includes Mel-cepstrum coefficients
and energy, and their first order differences. It is shown that as te
st utterance length increases, higher classification accuracy is achie
ved, Isolated word strings of 7-8 words uttered by the speaker results
in an accent classification rate of 93% among four different language
accents, A subjective listening test is also conducted in order to co
mpare human performance with computer algorithm performance in accent
discrimination. The results show that computer based accent classifica
tion consistently achieves superior performance over human listener re
sponses for classification. It is shown, however, that some listeners
are able to match algorithm performance for accent detection. Finally,
an experimental study is performed to investigate the influence of fo
reign accent on speech recognition algorithms. It is shown that traini
ng separate models for each accent rather than using a single model fo
r each word can improve recognition accuracy dramatically.