Phonetic Landmark Detection for Automatic Language Identification

This paper presents a method of augmenting shifted-delta cepstral coefficients (SDCCs) with the classification outputs of an array of support vector machines (SVMs) trained to detect a set of manner and place features on telephone speech. The SVM array allows for broad phoneme classification, and when this information is concatenated with SDCCs to form a hybrid feature vector for each acoustic frame, a set of Gaussian mixture models (GMMs) may be trained to perform automatic language identification (LID). The NTIMIT telephone band speech corpus was used to train the SVM-based distinctive feature recognizers, while the NIST callfriend telephone corpus was used for training and testing the rest of the system.