I may be missing something completely, but I take it that there are no python packages for windows testing currently. After Pip told me that it couldn’t find the correct package, I search PYPI to find that all files seemed to be for linux or OSX. Is there any talk of these packages being extended to windows, or is the only hope for windows to download the source and train the models? Thanks
So far we have no plan for Windows, but the model file itself is platform-independant, so you should be able to follow TensorFlow’s build instructions for Windows to produce the libraries and then build the bindings :).
Thanks for the quick response. I kinda figured that might be the case.
@Brandon_Beiler you can use WSL if you are in Windows 10. I wrote the instructions here: https://fotidim.com/deepspeech-on-windows-wsl-287cb27557d4
WSL approach will not work for serious windows application development. Does anyone has any experience of building this code (specifically client) for windows application.
@rao_jjn TensorFlow upstream has support for building on Windows, but so far we don’t do that, so it’s likely our code may not yet be ready for Windows. You’ll need to build TensorFlow for Windows if you want to hack on that, but we cannot help more at the moment.
Hi - We are also hoping for some Windows implementation at some point and are looking for developers who could help us build an application using deepspeech.
Also interested in getting this working on windows for https://github.com/getnamo/tensorflow-ue4
Looking at the DeepSpeech requirements.txt list, I suspect it will largely come down to being able to get all of these working in windows (or replacing them with an alternative with the same api)
- pandas - appears to install
- progressbar2 - appears to install
- python-utils - appears to install
- tensorflow == 1.4.0 - supported
- numpy - part of tensorflow
- matplotlib - appears to install
- scipy - appears to install
- paramiko >= 2.1 - appears to install (2.4)
- pysftp - appears to install
- sox - appears to install
- python_speech_features - appears to install
- pyxdg - appears to install
- bs4 - appears to install
- six - part of tensorflow
which do you think will be blockers?
This list of requirements is for training, I guess the topic was for inference, so the complexity of getting it on Windows is much lower than training on windows. Looking at your project, you might be interested in both, but I’m not sure how much training (e.g. using CUDA) is supported by tensorflow on windows?
You’re correct I’m interested in both.
GPU/cuDNN is already supported in tensorflow for windows (since 1.0) and in the plugin (check releases), so it should only boil down to ensuring the pre-requisites for deepspech work on windows. It appears on the surface that they all have a wheel for pip on windows, but I’ve yet to test the whole setup.
Great, patches are welcome if you get to something that works :).