Hi I would like to know the best ratio of WER(Word Error Rate) and EER (Equal Error Rate) for speech signals? Most of the studies just clarify on the results but I need to know the specific ratio to ensure that the results achieved within the standard acceptable rate. Thank you very much for any ideas on this.
Depending on speech quality and the amount of training data, these metrics can vary a lot.
For example, English read speech (librispeech open-scource data base) can be decoded with close to 4% WER, while for meeting speech with far-field microphone (AMI data base) you can get around 40%.
If your data set is standard, you should search for state-of-the-art papers using it.
I can also suggest looking at kaldi recipies for different data sets https://github.com/kaldi-asr/kaldi/tree/master/egs, where in each folder you have RESULTS file that summarizes WER for ASR and EER for speaker/language recognition.
You can also check some results here: https://github.com/syhw/wer_are_we
it Depends on your database speech quality also the good training of your approach may rasie this value.
- Badges
- Science topic