Atlas

Benchmarks

← All benchmarks

Earnings-22 (ESB test)

Multimodal · 2022-10-24

The Earnings-22 test split as segmented by the ESB benchmark and used by the Hugging Face Open ASR Leaderboard: 2,741 short segments (about five hours) of accented English earnings-call speech, not the corpus's full-length calls. Scores are corpus word error rate in percent after Whisper English normalization; lower is better.

Top models (lower is better)

ModelScore
LFM2.5-Audio-1.5B14.6
Loading Atlas data…