vernacularai/kaldi-train repository overview

Both CPU and GPU builds contain all Kaldi dependencies pre-installed that would be required during training. They both come along with pre-built Kaldi, SRILM toolkit, python-2.7.15 and python-3.6.5 which can be activated using pyenv.

GPU image was built on nvidia/kaldi:19.08-py3 and CPU on kaldiasr/kaldi:19-08. GPU build comes with pre-built CUDA-10.1 and support for NVIDIA GPUs.

Recommended way to spin off a GPU container is:

docker run --gpus all  -it -v /path/to/external/data/volume:/root/am-data -v /path/to/external/kaldi/training/repository:/workspace/kaldi --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 vernacularai/kaldi-train:gpu