For an internal project, I am looking for a Speech Recognition solution that fulfills the following criteria:
- It needs to be open-source and offline (i.e. not just using cloud based API's)
- It needs to be trainable and not just used for inference
- It needs to run in a Windows environment but could leverage on Cygwin or other "virtualization" options
- Given that it needs to be trainable, it also should support GPU/CUDA access
The packages which I have tried so far are Kaldi, DeepSpeech, DeepSpeech2, and CMU Sphinx
Kaldi and DS(2) are targeting Linux Using WSL (Windows Subsystem for Linux) would be ideal, except that it cannot utilize the GPU and thus is a no-go for training. Cygwin looked so far very promising, as it neatly exposes the GPU, but I am running into other problems compiling the packages (e.g. for Kaldi, I can't compile OpenFST on Cygwin) CMU Sphinx is the only package that works, but the quality has been very poor compared to Kaldi or DS(2)
I have spent quite a bit of time trying to make things work, but only identified the challenges so far.
If somebody managed to get a decent (Kaldi / DS(2) - type quality) system running under Windows then I would most welcome a recommendation.