Extending unsupervised phonetic segmentation with self-supervised speech embeddings

Abstract

Unsupervised phonetic segmentation provides a bottom-up alternative to forced alignment, enabling phonetic analyses without predefined segment labels and supporting a range of speech research scenarios where detailed phonetic annotation is unavailable or unnecessary. In previous work, I introduced a lightweight Praat-based implementation that segments speech into phone-sized units by computing framewise differences on traditional acoustic representations such as mel spectrograms and MFCCs, achieving competitive performance on benchmark data while remaining easy to install and use. In this paper, I extend that framework by incorporating self-supervised speech embeddings derived from recent deep learning models (e.g., wav2vec 2.0, HuBERT, WavLM, and Whisper) as alternative input representations. These embeddings are learned without phonetic supervision yet are known to encode contextualized phonetic structure and invariances that may not be readily captured by low-level acoustic features. This study extends prior work by comparing self-supervised embeddings with traditional acousticrepresentations using the same windowed difference measures and peak-based boundary detection procedures. Evaluation against manually annotated phonetic boundaries examines segmentation errors across phonetic boundary types, identifying conditions under which self-supervised embeddings exhibit different behavior while maintaining the simplicity of the original bottom-up approach.