Audio Feature Analysis for Precise Vocalic Segments Classification in English.

An approach to identifying the most meaningful Mel-Frequency Cepstral Coefficients representing selected allophones and vocalic segments for their classification is presented in the paper. For this purpose, experiments were carried out using algorithms such as Principal Component Analysis, Feature Importance, and Recursive Parameter Elimination. The data used were recordings made within the ALOFON corpus containing audio signal recorded employing 7 speakers who spoke English at the native or near-native speaker level withing a Standard Southern British English variety accent. The recordings were analyzed by specialists from the field of phonology in order to extract vocalic segments and selected allophones. Then parameterization was made using Mel Frequency Cepstral Coefficients, Delta MFCC, and Delta Delta MFCC. In the next stage, feature vectors were passed to the input of individual algorithms utilized to reduce the size of the vector by previously mentioned algorithms. The vectors prepared in this way have been used for classifying allophones and vocalic segments employing simple Artificial Neural Network (ANN) and Support Vector Machine (SVM). The classification results using both classifiers and methods applied for reducing the number of parameters were presented. The results of the reduction are also shown explicitly, by indicating parameters proven to be significant and those rejected by particular algorithms. Factors influencing the obtained results were discussed. Difficulties associated with obtaining the data set, its labeling, and research on allophones were also analyzed.