The RW-Voice EQ ASR evaluates open-source and proprietary English speech recognition models against human-curated references across four real-world conditions: accents, emotions, background audio (music & noise), and conversational dialogue.

Models are ranked by Average WER — the arithmetic mean across the four datasets (lower is better). We also report RTFx, the inverse real-time factor (higher is faster). Scoring follows HuggingFace's open_asr_leaderboard methodology — their text normalizer plus corpus-level WER — so numbers here are directly comparable to HF's.