There's something I don't understand: why would we remove all reference to Mozilla's Deepspeech project ? Has the project been terminated ? If so, could you link some articles about it ? And why Coqui ?

Hi that's a good point.
What initially got me into this was this issue here 3693
There has been some activity on this since yesterday. That is new to me. Not sure what is gonna come out of this. Nevertheless, I wanted to try coqui after reading this Best EVER 🐸 STT English model. The accuracy trained on common voice v7 feels improved while loading a much smaller model.

The accuracy trained on common voice v7 is really improved while loading a much smaller model.

@go-dockly Did you actually try it? This new model is just 1% better on real-life data and extremely slow (10 time slower than realtime).

Yes I tried it. How would I have created the PR?
The smaller model got subjectively better at understanding me yes.
Not perfect but definitely improved and running on gpu not too slow.
If it was faster and more accurate I'd be delighted. Give or take another year I guess.
I certainly like to try out the contributions you mentioned :)

Not perfect but definitely improved and running on gpu not too slow.

Haha, GPU is the case here. On CPU you have to wait for ages.

Did you ever try https://gitlab.com/Jaco-Assistant/Scribosermo ? It is much more accurate.

I certainly like to try out the contributions you mentioned :)

We are also uncertain on this just like you, waiting for Mozilla decision and very confused

We released unquantized versions of the acoustic model initially with the 1.0.0 release. This has since been fixed, performance should be equivalent or better than 0.9.3. If you're seeing something different I'd love to hear about it!

There's something I don't understand: why would we remove all reference to Mozilla's Deepspeech project ? Has the project been terminated ? If so, could you link some articles about it ? And why Coqui ?

Mozilla's lack of clear communication on this is above my pay grade, but as clearly visible from the activity in the repository, the DeepSpeech project is no longer maintained. I'm one of the main authors of DeepSpeech and am a co-founder at Coqui. We continue to build upon the project in the new repo, https://github.com/coqui-ai/STT, together with most of the community, who are now now hanging out in our Gitter room instead of Matrix, and discussions are now happening on GitHub instead of Discourse.

@reuben thanks for the insight and the continued improvements on DeepSpeech!

I certainly like to try out the contributions you mentioned :)

We are also uncertain on this just like you, waiting for Mozilla decision and very confused

@bobkleiner I am curious what would stop you guys from publishing those contributions in a separate repo?

Not perfect but definitely improved and running on gpu not too slow.

Haha, GPU is the case here. On CPU you have to wait for ages.

Did you ever try https://gitlab.com/Jaco-Assistant/Scribosermo ? It is much more accurate.

@bobkleiner Thanks for sharing I will give it a try.

What makes the recognition much better if you speak like the snippets in common voice.
I contribute voice samples often to that project and know that many sentence snippets are in relation to people and places.
So phrases "he lived in iowa" are well understood whereas terms like "alpha echo bravo charlie" or "punct comma enter" are just not in it's grasp. It would also be beneficial if the source of those sentences was more casual in the common voice app by for instance using subtitle snippets of movies.

@reuben do you consider changing the sentence source in the common voice app realistically doable? What is the future of the common voice project? Do you think Mozilla might shut it down?

I no longer have any influence on the Common Voice project, so I can't comment on their project direction. It's also possible that the language model text source in our release models doesn't include such phrases as "alpha echo bravo charlie" and so on. Experimenting on the text side of things is cheaper and faster than augmenting the labeled speech data and re-training or fine tuning the acoustic model, so I encourage trying that first. And please report results! :D

Thanks for all the clarifications.

I've decided to create a new repo specifically for Coqui since I'm not comfortable using this one to do so. I've started renaming files and packages so that Coqui is mentioned instead of Deepspeech.

@go-dockly could you:

1. Create a new PR in this new repo with almost the same changes you submitted here (minus @reuben comments + there shouldn't be any mention of Deepspeech or Mozilla anymore)?
2. Use this PR to only indicate at the top of the README that a new repo has been created for Coqui bindings?

Cheers

let's do it :)