sanchit-gandhi

/

distil-whisper-large-v3-de-kd

Like 9

Hi, I tested the model myself and got a bit confused with the reported performance, in the model card the eval set WER is actually the validation set WER (not the test set which I assume is used for evaluation).
The WER matches for the validation set, in case someone is interested in the results for the test set, these are my results:

database: common_voice_german
test set:
WER (unnorm.): 9.09%
WER(norm.): 7.36%
validation set:
WER (unnorm.): 8.13%
WER(norm.): 6.29%