Phoneme- and Word-Level Metrics Using Self-Supervised Speech Representations for Forced Alignment Evaluation
The paper proposes reference-free alignment scoring that does not need manually annotated timestamps.
It introduces two corpus-level metrics, PCMI for phoneme-label agreement with SSL-derived clusters and WACS for acoustic consistency across repeated words. The authors test both under random and systematic alignment perturbations and report that the scores fall as alignments degrade. They evaluate across 85 FLEURS languages, validate against manual alignments in 45 DoReCo languages, and include two phonologically complex low-resource languages. The metrics are released as an open-source Python package. ArXiv · AI/CL/LG's note
It introduces two corpus-level metrics, PCMI for phoneme-label agreement with SSL-derived clusters and WACS for acoustic consistency across repeated words. The authors test both under random and systematic alignment perturbations and report that the scores fall as alignments degrade. They evaluate across 85 FLEURS languages, validate against manual alignments in 45 DoReCo languages, and include two phonologically complex low-resource languages. The metrics are released as an open-source Python package. ArXiv · AI/CL/LG's note
score 4