1*a58d3d2aSXin Li# LPCNet 2*a58d3d2aSXin Li 3*a58d3d2aSXin LiIncomplete pytorch implementation of LPCNet 4*a58d3d2aSXin Li 5*a58d3d2aSXin Li## Data preparation 6*a58d3d2aSXin LiFor data preparation use dump_data in github.com/xiph/LPCNet. To turn this into 7*a58d3d2aSXin Lia training dataset, copy data and feature file to a folder and run 8*a58d3d2aSXin Li 9*a58d3d2aSXin Lipython add_dataset_config.py my_dataset_folder 10*a58d3d2aSXin Li 11*a58d3d2aSXin Li 12*a58d3d2aSXin Li## Training 13*a58d3d2aSXin LiTo train a model, create and adjust a setup file, e.g. with 14*a58d3d2aSXin Li 15*a58d3d2aSXin Lipython make_default_setup.py my_setup.yml --path2dataset my_dataset_folder 16*a58d3d2aSXin Li 17*a58d3d2aSXin LiThen simply run 18*a58d3d2aSXin Li 19*a58d3d2aSXin Lipython train_lpcnet.py my_setup.yml my_output 20*a58d3d2aSXin Li 21*a58d3d2aSXin Li## Inference 22*a58d3d2aSXin LiCreate feature file with dump_data from github.com/xiph/LPCNet. Then run e.g. 23*a58d3d2aSXin Li 24*a58d3d2aSXin Lipython test_lpcnet.py features.f32 my_output/checkpoints/checkpoint_ep_10.pth out.wav 25*a58d3d2aSXin Li 26*a58d3d2aSXin LiInference runs on CPU and takes usually between 3 and 20 seconds per generated second of audio, 27*a58d3d2aSXin Lidepending on the CPU. 28