CNN-Based Phonetic Segmentation Refinement with a Cross-Speaker Setup

Abstract

This work proposes a method to improve the performance of automatic phonetic alignment of speech data. The method uses a deep convolutional neural network (CNN) trained on a combination of acoustic features extracted from labeled data to fine tune the position of each boundary within a fixed-size window around the original boundary position. The proposed method is robust to speaker identity, which means that a system trained with enough labeled data can be used to fine tune alignment on any speech file, regardless of speaker identity. With an absolute gain between 20% and 33% in cross speaker scenario, our results demonstrate the applicability of deep learning for this task.