New method enables speech evaluation in low-resource languages by eliminating the need for precise phoneme timing alignment

Researchers propose a solution to adapt weakly-supervised ASR models for speech evaluation tasks that traditionally require phoneme-level time boundaries

The approach uses word-level instead of phoneme-level speaking rate and duration metrics to work around limitations of frame-asynchronous models

Phoneme posteriors are extracted by mapping ASR hypotheses to phoneme confusion networks rather than direct phoneme recognition

A cross-attention architecture combines phoneme and frame-level features, eliminating the need for phoneme time alignment

The method achieves comparable performance to standard frame-synchronous features on English speech while enabling expansion to low-resource languages