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

• Published in: Computational Processing of the Portuguese Language Vol. 11122; pp. 448 - 456
• Main Authors: Cuozzo, Luis Gustavo D., Silva, Diego Augusto, Neto, Mario Uliani, Simões, Flávio Olmos, Nagle, Edson Jose
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2018; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Deep neural networks; Forced alignment; Phonetic segmentation refinement
• ISBN: 9783319997216; 3319997211
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-319-99722-3_45

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.