1*a58d3d2aSXin Li# Opus Speech Coding Enhancement 2*a58d3d2aSXin Li 3*a58d3d2aSXin LiThis folder hosts models for enhancing Opus SILK. 4*a58d3d2aSXin Li 5*a58d3d2aSXin Li## Environment setup 6*a58d3d2aSXin LiThe code is tested with python 3.11. Conda setup is done via 7*a58d3d2aSXin Li 8*a58d3d2aSXin Li 9*a58d3d2aSXin Li`conda create -n osce python=3.11` 10*a58d3d2aSXin Li 11*a58d3d2aSXin Li`conda activate osce` 12*a58d3d2aSXin Li 13*a58d3d2aSXin Li`python -m pip install -r requirements.txt` 14*a58d3d2aSXin Li 15*a58d3d2aSXin Li 16*a58d3d2aSXin Li## Generating training data 17*a58d3d2aSXin LiFirst step is to convert all training items to 16 kHz and 16 bit pcm and then concatenate them. A convenient way to do this is to create a file list and then run 18*a58d3d2aSXin Li 19*a58d3d2aSXin Li`python scripts/concatenator.py filelist 16000 dataset/clean.s16 --db_min -40 --db_max 0` 20*a58d3d2aSXin Li 21*a58d3d2aSXin Liwhich on top provides some random scaling. 22*a58d3d2aSXin Li 23*a58d3d2aSXin LiSecond step is to run a patched version of opus_demo in the dataset folder, which will produce the coded output and add feature files. To build the patched opus_demo binary, check out the exp-neural-silk-enhancement branch and build opus_demo the usual way. Then run 24*a58d3d2aSXin Li 25*a58d3d2aSXin Li`cd dataset && <path_to_patched_opus_demo>/opus_demo voip 16000 1 9000 -silk_random_switching 249 clean.s16 coded.s16 ` 26*a58d3d2aSXin Li 27*a58d3d2aSXin LiThe argument to -silk_random_switching specifies the number of frames after which parameters are switched randomly. 28*a58d3d2aSXin Li 29*a58d3d2aSXin Li## Regression loss based training 30*a58d3d2aSXin LiCreate a default setup for LACE or NoLACE via 31*a58d3d2aSXin Li 32*a58d3d2aSXin Li`python make_default_setup.py model.yml --model lace/nolace --path2dataset <path2dataset>` 33*a58d3d2aSXin Li 34*a58d3d2aSXin LiThen run 35*a58d3d2aSXin Li 36*a58d3d2aSXin Li`python train_model.py model.yml <output folder> --no-redirect` 37*a58d3d2aSXin Li 38*a58d3d2aSXin Lifor running the training script in foreground or 39*a58d3d2aSXin Li 40*a58d3d2aSXin Li`nohup python train_model.py model.yml <output folder> &` 41*a58d3d2aSXin Li 42*a58d3d2aSXin Lito run it in background. In the latter case the output is written to `<output folder>/out.txt`. 43*a58d3d2aSXin Li 44*a58d3d2aSXin Li## Adversarial training (NoLACE only) 45*a58d3d2aSXin LiCreate a default setup for NoLACE via 46*a58d3d2aSXin Li 47*a58d3d2aSXin Li`python make_default_setup.py nolace_adv.yml --model nolace --adversarial --path2dataset <path2dataset>` 48*a58d3d2aSXin Li 49*a58d3d2aSXin LiThen run 50*a58d3d2aSXin Li 51*a58d3d2aSXin Li`python adv_train_model.py nolace_adv.yml <output folder> --no-redirect` 52*a58d3d2aSXin Li 53*a58d3d2aSXin Lifor running the training script in foreground or 54*a58d3d2aSXin Li 55*a58d3d2aSXin Li`nohup python adv_train_model.py nolace_adv.yml <output folder> &` 56*a58d3d2aSXin Li 57*a58d3d2aSXin Lito run it in background. In the latter case the output is written to `<output folder>/out.txt`. 58*a58d3d2aSXin Li 59*a58d3d2aSXin Li## Inference 60*a58d3d2aSXin LiGenerating inference data is analogous to generating training data. Given an item 'item1.wav' run 61*a58d3d2aSXin Li`mkdir item1.se && sox item1.wav -r 16000 -e signed-integer -b 16 item1.raw && cd item1.se && <path_to_patched_opus_demo>/opus_demo voip 16000 1 <bitrate> ../item1.raw noisy.s16` 62*a58d3d2aSXin Li 63*a58d3d2aSXin LiThe folder item1.se then serves as input for the test_model.py script or for the --testdata argument of train_model.py resp. adv_train_model.py 64*a58d3d2aSXin Li 65*a58d3d2aSXin LiCheckpoints of pre-trained models are located here: https://media.xiph.org/lpcnet/models/lace-20231019.tar.gz