Koel Labs Research

English Speech-to-IPA Leaderboard

Compare English phonetic transcription models for speech-to-IPA accuracy across accents and speech impairments.

Evaluation, compute & contributions

Evaluation

  • PER (Phoneme Error Rate): Levenshtein distance between predicted and reference phoneme sequences.
  • FER (Feature Error Rate): Edit distance weighted by phonetic features from panphon.

Models are evaluated across native, non-native, and impaired English speech. Read the metrics guide or how this leaderboard was built.

Compute and contributions

This leaderboard runs on the basic Hugging Face Spaces plan (16GB RAM, 2 vCPUs). Evaluations can take several hours; please submit each model only once.

Read the README, open a Space discussion, or contact info@koellabs.com.

Sort by average
Loading Leaderboard...