Mining speech signal patterns for robust speaker variability classification

A speaker identification framework that combines both high- and low-level features, for state-of-the-art variability analysis and classification is proposed in this paper. The proposed framework introduces a workable solution that targets robust speaker variability classification using speech samples recorded in suboptimal conditions. A translated Ibibio (New Benue-Congo, Nigeria) version of “The Tiger and the Mouse”—a prosodically balanced corpus that demonstrates the prosody of read-aloud English was used in this study with speech samples obtained from 50 participants (25 males and 25 females). Identity-vectors (i-vectors) or low dimensional signal patterns were extracted and used as baselines for investigating speakers’ variability patterns across various classifiers (Decision Tree: DT, Support Vector Machine: SVM, k-Nearest Neighbour: k-NN, and Deep Neural Network: DNN) and kernels. Treatment of the baselines with high level features (speech duration, F0, intensity) was also experimented for word, syllable, and phoneme units. Results obtained revealed that DTs and some variants of SVM gave high classification accuracies (above 70%). Hence, the hypothesis of universal Gaussianity appears inexact, as the linear predictor that is optimal in the mean square error sense may not hold for Ibibio. Further treatments of the baselines with Linear Discriminant Analysis (LDA) and Cosine Distant Scoring (CDS) yielded very poor classification results, except for the k-NN and Gaussian SVM classifiers which performed well for the LDA treated baselines.