coqui-ai / stt-models Goto Github PK

Can't put email and can't download models.

Please enter your email for model download - didnt work

The documentation (here) for the English yesno model says that the model was trained on the Common Voice Target Segments Corpus.

However this corpus does not seem to be available online. The documentation should provide a link to the location of this dataset.

Hi,
i just tried the german model from Aashish and run into a segmentation fault.
On the overview page it is marked for Coqui STT v1.0.0 but on the next page it is stated,
that it is only for STT version: v0.9.3.
It´s kind of confusing
Kind regards,
Nils

It seems https://coqui.ai/models is not reachable anymore.
Will you host the models somewhere else?

Some models have >85% WER and others have <15% WER but it is hard to know which models are reasonably usable without clicking on each one individually. It would be helpful to have WER as a sortable column on the Models page of the site.

Sometimes WER information is hard to find even on the model's card page. Some models direct you to other pages to view that information and the ones that do display it show it in different formats. It would be good if there was some form of standardization on how to report WER within the model cards.

The alphabet files from Jaco models are inconsistent with the output of the models at runtime. It has been observed that the Jaco Spanish model can produce accented vowels, but the alphabet file does not include them. The alphabet file should be confirmed and uploaded to the zoo for language model generation.

TFModelState::init and TFLiteModelState::init can be modified to print out the loaded alphabet used to train the model here: https://github.com/coqui-ai/STT/blob/653ce25a7ce5bd6cbb564416d847d8afcd5c5e8c/native_client/tfmodelstate.cc#L120