Spaces:

---

hf-audio

/

open_asr_leaderboard

like 1.07k

Running on CPU Upgrade

Hey guys, the dialect mismatch in the datasets you posted are going to cause some issues down the line. May make life easier to specify them by dialect so the WER performance is more interpretable.

• Portuguese
  -> Fleurs is from Brazilian dialect.
  -> MLS is mix of Brazilian and European. But the texts are going to be biased towards European dialect but pronounciation will be heavily Brazilian.

• Spanish
  -> Fleurs is LA
  -> MLS will be biased towards European dialect in writing but LA dialect in pronunciation. (Don Quixote read by someone from Mexico, for instance).