Spaces:

huggingface

/

hf-speech-bench

like 81

Running

In the literature, WER results are usually quoted to 1 d.p. (c.f. https://arxiv.org/abs/2006.11477). The Speech Bench contains a mixture of formats for quoted results, to both 1 and 2 d.p. Given that official results are quoted to 1 d.p., should we round WER figures to 1 d.p.? Or do we risk a loss of granularity with this precision (i.e. three models with WERs of 1.95%, 2.01% and 2.04% will all be rounded to 2.0%).