Flexible and Interpretable Accent Distance Measurements

Details

Measuring accent differences typically forces a tradeoff: phonetics researchers use vowel formants from controlled paired recordings for interpretability, but data collection is expensive and may not capture connected speech; TTS researchers use accent classification embeddings trainable from any audio, but the results resist interpretation. This paper proposes articulatory representations—created via articulatory inversion from standard recordings—as an interpretable basis for accent comparison. Optimal transport provides the distance metric, enabling comparisons across recordings of different lengths and types without requiring paired controlled stimuli. The approach bridges both research methodologies rather than replacing either. The paper is light on quantitative benchmarks, making it difficult to assess performance against existing methods. From a practical speech technology standpoint, interpretable accent distance has applications in TTS accent control, dialect research, and language learning tools—though broader adoption will depend on concrete comparisons with embedding-based baselines.