Unsupervised speech segmentation: an analysis of the hypothesized phone boundaries

Area of Science:

• Speech Processing
• Computational Linguistics
• Machine Learning

Background:

• Unsupervised automatic phone segmentation methods often yield similar performance in boundary detection.
• These methods do not perfectly replicate human-created reference transcriptions.

Purpose of the Study:

• Investigate fundamental issues in unsupervised segmentation algorithms.
• Compare acoustic-only segmentation with human reference transcriptions.

Main Methods:

• Analyzed an unsupervised speech segmentation method using acoustic change detection.
• Compared hypothesized boundaries with human-annotated segment boundaries.
• Performed statistical analyses on segmentation errors.

Main Results:

• Acoustic change is a reliable indicator for segment boundaries, with over two-thirds accuracy.
• Errors correlate with segment duration, similar segment sequences, and dynamic phones.
• Current one-stage methods need enhancement into two-stage approaches.

Conclusions:

• Enhancing unsupervised segmentation requires integrating bottom-up and top-down information.
• Two-stage methods can improve accuracy while maintaining flexibility and language independence.