A Hierarchical Classification Framework for Phonemes and Broad Phonetic Groups (BPGs): a Discriminative Template-Based Approach

In this paper, a novel framework to phone or phoneme classification is presented. The framework combines discriminative classification approach to the traditional HMM framework. Unlike the traditional HMM approach to phoneme recognition, here all phones are modeled by one HMM. However, instead of using generative models (e.g., GMMs or codebooks), this framework employs a discriminative classifier to predict the state probabilities and finds the optimal state sequence to obtain a time-alignment function between the acoustic feature vector sequence and the state sequence. For each state Si, the corresponding feature vectors are averaged resulting in a single feature vector that represents the i-th vector of the block. All feature vectors of the block are then concatenated to a single feature vector to represent a phone unit, which is used as a feature vector for a phone classifier. The phone classifier is hierarchical is the sense that the broad phonetic groups (BPGs) are classifier followed by the phonemes belonging to those classes. Validated by the TIMIT database, the proposed framework with MFCCs has comparable performance to related phoneme classification algorithms, but with flexibility to account for duration and other features such as articulatory features. We also observe that the framework gives promising results for BPG classification.