||Presently, speech recognition is gaining worldwide popularity in applications like Google Voice, speech-to-text reporter (speech-to-text transcription, video captioning, real-time transcriptions), hands-free computing, and video games. Research has been done for several years and many speech recognizers have been built. However, most of the speech recognizers fail to recognize the speech accurately. Consider the well-known application of Google Voice, which aids in users search of the web using voice. Though Google Voice does a good job in transcribing the spoken words, it does not accurately recognize the words spoken with different accents. With the fact that several accents are evolving around the world, it is essential to train the speech recognizer to recognize accented speech. Accent classification is defined as the problem of classifying the accents in a given language. This thesis explores various methods to identify the accents. We introduce a new concept of clustering windows of a speech signal and learn a distance metric using specific distance measure over phonetic strings to classify the accents. A language structure is incorporated to learn this distance metric. We also show how kernel approximation algorithms help in learning a distance metric.