Intelligent Speech Features Mining for Robust Synthesis System Evaluation

摘要

Speech synthesis evaluation involves the analytical description of useful features, sufficient to assess the performance of a speech synthesis system. Its primary focus is to determine the degree of semblance of synthetic voice to a natural or human voice. The task of evaluation is usually driven by two methods: the subjective and objective methods, which have indeed become a regular standard for evaluating voice quality, but are mostly challenged by high speech variability as well as human discernment errors. Machine learning (ML) techniques have proven to be successful in the determination and enhancement of speech quality. Hence, this contribution utilizes both supervised and unsupervised ML tools to recognize and classify speech quality classes. Data were collected from a listening test (experiment) and the speech quality assessed by domain experts for naturalness, intelligibility, comprehensibility, as well as, tone, vowel and consonant correctness. During the pre-processing stage, a Principal Component Analysis (PCA) identified 4 principal components (intelligibility, naturalness, comprehensibility and tone) - accounting for 76.79% variability in the dataset. An unsupervised visualization using self organizing map (SOM), then discovered five distinct target clusters with high densities of instances, and showed modest correlation between significant input factors. A Pattern recognition using deep neural network (DNN), produced a confusion matrix with an overall performance accuracy of 93.1%, thus signifying an excellent classification system.