Phoneme- and Word-Level Metrics Using Self-Supervised Speech Representations for Forced Alignment Evaluation