I would like to use the pre-trained model that was released with version 0.4.1, with the 0.5.0 alpha build of Deep Speech. How can I do so? I am getting a compatibility error.
Any help appreciated!
Nobody has tried that yet, so I’m afraid you’re breaking new ground!
1 Like
Wow! Really? How are the new alpha versions tested without voice models?
By training small models for the sole purpose of testing the code 
The alpha versions are really not meant for general use, they represent partial work towards a new stable release. That said, it may be possible to re-export the 0.4.1 checkpoints with the current code to get a working model, that’s what I was referring to when I said nobody has tried it. If you can get it to work, feel free to submit a PR. I’m sure we’ll get to it eventually before the next release, even if just to confirm if it’s possible.
Awesome thanks for the info. I am on Windows and (I’m not sure about this but) I believe that Windows integration is new with the alpha builds. I tried to re-export the 0.4.1 checkpoints with the current code to get a working model, but have had some difficulties that I will try to work out as the windows version doesn’t use DeepSpeech.py, making these instructions: https://github.com/mozilla/DeepSpeech#continuing-training-from-a-release-model obsolete.
In order to re-export you’ll need to run the training code, which only works on Linux and macOS. Currently our Windows support is for the clients only.
– reuben
I’ll see if I can get it to run on my raspberry pi, Thanks for the help!
I’m also very eager to get a pre trained model to be used with the Windows client, which is only available in 0.5.0. The only one I see on the releases page was done for 0.4.1, and the client rejects it.
For those of you with more experience on this project, would you recommend I:
1. Wait for the release of 0.5.0 (I can wait a week, but probably not a month or more)
2. Try to get 0.4.1 compiling on Windows. I see a guide, but I also see a lot of changes around getting Windows to compile (pathing issues and such), and worry that this may be a great deal of work. I can spend a day or two, but probably not a week.
We are really hoping to use DeepSpeech on a VR experience for a conversation engine, that we’ll be releasing in June.
Hudson, have you made any progress here? I’d be so happy if you could share a 0.5.0 model if you have one.
Thegiffman, today I am setting up a Linux machine to attempt and convert the pre-trained model to the alpha version mentioned. I am not sure how long the checkpoint process will take though. I will update as I progress!
I’d go for option 2 if you can’t wait a month or more. There’s no v0.5 model right now that will perform well, as we changed the feature computation code since v0.4.1. Uplifting the changes to v0.4.1 shouldn’t be too hard, depends on how experienced you are with C++ and our build system I suppose. I can’t tell you if it’s going to take more than a day or two.
Given what Reuben said, I now see that there is no point in attempting to convert the model as the alpha wouldn’t perform well anyways.
Well, I’ve got zero experience with your build system, but I’ve used Visual Studio a decent amount to compile C++. I’ll give it my best effort! If I run into trouble, I may start a thread. You all seem like kind souls, who would be willing to offer a tip or two to someone giving a good faith effort.
Thanks Reuben, for steering us in the right direction.
@hudson, you’ll want to check out my other thread here. @carlfm01 just saved me a day at least of work by sharing his Windows build of 0.4.1. Maybe it will come in handy for you as well?
Oh sick, definitely will check that out when I get time. Thanks for the link!