Department or Program
Cognitive and Linguistic Sciences
Primary Wellesley Thesis Advisor
As speech recognition and intelligent systems are more prevalent in society today, we need to account for the variety of accents in spoken language. An important step involves identifying the type of accent given a sample of speech. For this thesis, we have coded machine learning algorithms to classify accents from foreign-accented English. Given a data set of 4925 phone calls that span 23 different accents, we have trained Gaussian Mixture Models for each accent with two main methods. The text-independent classifier assumes that we took sound features without knowing the transcriptions of the speech, while the text-dependent classifier relies on transcriptions in order to align each phoneme (or sound unit, e.g. /AH/ and /K/) to its utterance in the data. We acquired these transcriptions by releasing tasks via Amazon Mechanical Turk for the following 7 accents: Arabic, Czech, French, Hindi, Indonesian, Korean, and Mandarin. Upon evaluation, we found that a 7-way accent identification task achieved an accuracy rate of 41.38% for the text-independent classifier and 45.12% for the text-dependent classifier.