Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Number of Words in Test Set #87

Open
xiaoxiao26-zz opened this issue Aug 20, 2021 · 4 comments
Open

Number of Words in Test Set #87

xiaoxiao26-zz opened this issue Aug 20, 2021 · 4 comments
Labels
documentation Improvements or additions to documentation

Comments

@xiaoxiao26-zz
Copy link

xiaoxiao26-zz commented Aug 20, 2021

Hi, is there an official number for the final number of words in the test set for scoring?

WeNet results say there are 19928 sentences and 390656 words: https://github.com/wenet-e2e/wenet/tree/main/examples/gigaspeech/s0

Kaldi results say there are 390721 words: https://github.com/kaldi-asr/kaldi/blob/master/egs/gigaspeech/s5/RESULTS

I am using my own toolkit, and after removing the noise/punc tags and running the provided scoring script, I get 19930 sentences and 390744 words. Note I do match WeNet's number of words/sentences on the dev set.

Although these are small differences, I guess this still matters for final reporting. It would be super helpful if:

  1. There was an official number of sentences/words, Kaldi vs WeNet seems to suggest there are inconsistencies.
  2. To aid with debugging, a file listing the final references post filtering + normalization for each segment would be helpful. This way people can easily check if they are evaluating on the exact same reference, and if not where the filtering/normalization went wrong.

Thank you!

@chenguoguo
Copy link
Collaborator

Thanks, we will get back to this.

@dophist
Copy link
Collaborator

dophist commented Aug 22, 2021

Another reference of GigaSpeech EvaluationSets numbers can also be found here in Espnet recipe, it is 19930 utts & 390744 words.

Guoguo has assigned @chaisz19 to confirm these numbers, something that I can think of:

  1. some downstream toolkits might filter long utterances(say longer than 20 sec). GigaSpeech utterances are between [0.5, 20] seconds, but float-point rounding error(around 20) may be a source of inconsistency during utterance filtering.
  2. After text post-processing, some utterances might have empty reference(say, an utterance with raw reference 'ah um'), I'm not sure Kaldi's compute-wer and sclite handle denominator numbers exactly the same way in this situation.

Given the large scale of GigaSpeech evaluation sets, this shouldn't have any significant impact on final WER, but it's good though that we can confirm on these numbers. So let's wait for Shuzhou's confirmation.

By the way, @xiaoxiao26 it would be great if you can provide the md5 of your local GigaSpeech.json, so we can make sure the investigation is on the right basis.

@xiaoxiao26-zz
Copy link
Author

Thanks @dophist, good to know that the Espnet recipe exactly matches my numbers. md5sum returns the following:

19c777dc296ff3eb714bc677a80620a3 data/GigaSpeech.json

@dophist dophist added the documentation Improvements or additions to documentation label Jan 6, 2022
@dophist
Copy link
Collaborator

dophist commented Jan 6, 2022

Leave this thread open for documentation purpose. As long as the MD5 is consistent with above, everything should be fine.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
documentation Improvements or additions to documentation
Projects
None yet
Development

No branches or pull requests

3 participants