Phone segmentation and recognition are closely related, but modern speech systems typically handle them in separate stages. A new arXiv paper challenges this separation, arguing that the phonetic information needed for both tasks is already present in the representations learned by self-supervised speech models.
The authors propose a method called phonological activation mapping, which aims to extract that latent structure directly. Rather than training separate models for segmentation and recognition, their approach would use the same underlying representations to identify phone boundaries and phone identities simultaneously.
Because this is a single arXiv abstract, the details are limited and the claims are not yet independently verified. The paper does not report experimental results in the abstract, so it remains an argument for a different modeling strategy rather than a demonstrated improvement. Still, if the latent-structure hypothesis holds, it could simplify speech processing pipelines and make better use of existing self-supervised features.