Add pretrained models #435
Comments
|
We're planning on adding links to some pre-trained models as soon as we're satisfied with their word error rate (WER). We're targeting a WER of at max 10% on the TED test set and hope that we can, within the next few weeks, release a model. |
|
@kdavis-mozilla is there any update on this line of work? |
|
@gvoysey I wish I could give you the models right now. But I can't. Sorry. The WER still needs to be tuned. We're still waiting on our new hardware which will allow us to tune hyperparameters with a quicker turn around. But it's not here yet. Unfortunately the ETA for the hardware looks to be about 4 weeks out now. Then, once we have the hardware, we'll still need to tune for about a week or two. Again, sorry. |
|
@kdavis-mozilla no worries. I have been thrashing some GPUs myself (on librivox), hopefully they'll finish soon! |
|
I'm curious how large a trained model from a large dataset like TED is on disk, as a rough estimate? |
|
I have a trained librivox model which is ~800 MB. I've found deepspeech2 implementations on other systems (torch) that are ~600 MB. These facts are anecdata. :) |
|
You know how long will it take to train with the TED dataset, if I have a GTX 1080? |
|
@phasnox We use 4 Titan X's and it takes about 5 days. So, I'd guess on the order of 20 days. |
|
@kdavis-mozilla maxwell Titan X, Titan X (Pascal), or Titan Xp? I hate that we need the distinction... |
|
@getnamo Titan X (Pascal)
But I don't think NVIDIA even tried on this one. |
|
and off by one errors 😇
…On Tue, May 16, 2017 at 02:46 Kelly Davis ***@***.***> wrote:
@getnamo <https://github.com/getnamo> Titan X (Pascal)
There are only two hard things in Computer Science:
cache invalidation and naming things.
-- Phil Karlton
But I don't think NVIDIA even tried on this one.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#435 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ADeR7_MjPvbL9okWh4HHQxcYGiyMKSnrks5r6TywgaJpZM4MaXNN>
.
|
|
Any news for links to trained models? |
|
@phasnox 20 days for how many epochs? WER achieved? |
|
@striki70 yehp WER achieved would be nice. But I have'nt try yet, I need to get a better cooling system. |
|
@kdavis-mozilla just checking in on any updates you have on trained models. |
|
Just got the hardware and installed it last week. Now we're working through a few OOM issues which appear in the cluster setting and not in the single node setting. So, it's a work in progress. |
|
Waiting for it :) Thanks. |
|
Are there any updates regarding the pretrained models? |
|
We're training a bunch of models and getting ready to share them. It's coming soon! :) |
|
Are they ready now? |
|
We'll comment here when they're available. |
|
I test TED dataset with Google API using this tool autosub with little modification. Google API achieved 27.3162% WER. Is 10% WER aiming too high? |
|
That's off topic for this issue. Please join our IRC channel for questions and discussions. I'm locking this issue to try and get people to stick to the proper communication channels. |
|
We have a first release of an American English model available in our releases page: https://github.com/mozilla/DeepSpeech/releases/latest Please check it out and experiment, we're excited to see what you can do with it. We also set up discussion forums on Discourse, check the release notes for links. |
|
@reuben We surely appreciate the work and effort you have taken in, to provide us the model. It works amazing. It would be much more helpful, if you could provide us the checkpoints of the above model. |
|
@reuben @kdavis-mozilla and others, thanks for the great work (amazing, really !) |
|
@nicolaspanel Thanks, and yes we want to share models for as many languages as we can! |
|
@kdavis-mozilla great 👍 do you have any ETA (no mentioned in currents projects)? |
|
@nicolaspanel For other languages the timing depends upon the rate at which Common Voice collects data for the language. For example, a couple of weeks ago data donation started for French. The faster Common Voice collects data for a particular language, e.g. French, the faster we can bring you a model for that language. So to a very large extent the timings will be determined be the community of Common Voice. |
|
@nicolaspanel For french we also deeply need to diversify our sources. If you're interested, you can find informations on https://github.com/mozfr/besogne/wiki/Common-Voice-Fr |
|
Are there any news on next trained model release? Thanks a lot, your work is amazing! |
|
No real updates yet, but as we just started collecting in other languages we need to give it a bit more time before we have enough data to train on. |
|
PS: And thanks for the compliment! |
|
Are you working towards increasing the dataset or trying new networks? Or both maybe? |
|
Right now trying to get more data through Common Voice |
|
Have you thought about data mining movies/series with their subtitles? |
|
Licensing issues prevent this. |
|
Oh I understand, thanks again for your time, Keep up the great work! |
|
Is there any pretrained model of indian english pls send me if it exists |
|
@akshat9425 A couple of us are working towards achieving something on similar lines. We could explore the possibility of collaborating on the same. Write to me bak0@protonmail.com |
|
This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs. |
Is it possible to add links to some pretrained models? I would like to test the performance on some real world speech, but could not find any reference to these in the docs. Thanks!
The text was updated successfully, but these errors were encountered: