Spaces:

---

hf-audio

/

open_asr_leaderboard

like 1.36k

Running on CPU Upgrade

Hello Hugging Face Audio Team,

Thank you for all your great work! I had a quick question regarding the evaluation metrics for a couple of models.

Could you help clarify why there's a significant difference in WER on the librispeech-clean dataset between facebook/wav2vec2-large (12.81 WER) and speechbrain/asr-wav2vec2-librispeech (1.77 WER)?
My hypothesis is that the Facebook model might only be using a linear CTC head, while the SpeechBrain model uses language model head. Could you confirm if this is the case?
Thanks so much.

Kind Regards,
Goksenin Yuksel