Awesome

micro-TCN

| Paper | Demo | Plugin |

Efficient neural networks for real-time modeling of analog dynamic range compression.

Setup

Install the requirements.

python3 -m venv env/
source env/bin/activate
pip install -r requirements.txt

Then install auraloss.

pip install git+https://github.com/csteinmetz1/auraloss

Pre-trained models

You can download the pre-trained models here. Then unzip as below.

mkdir lightning_logs
mv models.zip lightning_logs/
cd lightning_logs/
unzip models.zip 

Use the compy.py script in order to process audio files. Below is an example of how to run the TCN-300-C pre-trained model on GPU. This will process all the files in the audio/ directory with the limit mode engaged and a peak reduction of 42.

python comp.py -i audio/ --limit 1 --peak_red 42 --gpu

If you want to hear the output of a different model, you can pass the --model_id flag. To view the available pre-trained models (once you have downloaded them) run the following.

python comp.py --list_models

Found 13 models in ./lightning_logs/bulk
1-uTCN-300__causal__4-10-13__fraction-0.01-bs32
10-LSTM-32__1-32__fraction-1.0-bs32
11-uTCN-300__causal__3-60-5__fraction-1.0-bs32
13-uTCN-300__noncausal__30-2-15__fraction-1.0-bs32
14-uTCN-324-16__noncausal__10-2-15__fraction-1.0-bs32
2-uTCN-100__causal__4-10-5__fraction-1.0-bs32
3-uTCN-300__causal__4-10-13__fraction-1.0-bs32
4-uTCN-1000__causal__5-10-5__fraction-1.0-bs32
5-uTCN-100__noncausal__4-10-5__fraction-1.0-bs32
6-uTCN-300__noncausal__4-10-13__fraction-1.0-bs32
7-uTCN-1000__noncausal__5-10-5__fraction-1.0-bs32
8-TCN-300__noncausal__10-2-15__fraction-1.0-bs32
9-uTCN-300__causal__4-10-13__fraction-0.1-bs32

We also provide versions of the pre-trained models that have been converted to TorchScript for use in C++ here.

Evaluation

You will first need to download the SignalTrain dataset (~20GB) as well as the pre-trained models above. With this, you can then run the same evaluation pipeline used for reporting the metrics in the paper. If you would like to do this on GPU, perform the following command.

python test.py \
--root_dir /path/to/SignalTrain_LA2A_Dataset_1.1 \
--half \
--preload \
--eval_subset test \
--save_dir test_audio \

In this case, not only will the metrics be printed to terminal, we will also save out all of the processed audio from the test set to disk in the test_audio/ directory. If you would like to run the tests across the entire dataset you can specific a different string after the --eval_subset flag, as either train, val, or full.

Training

If would like to re-train the models in the paper, you can run the training script which will train all the models one by one.

python train.py \ 
--root_dir /path/to/SignalTrain_LA2A_Dataset_1.1 \
--precision 16 \
--preload \
--gpus 1 \

Plugin

We provide plugin builds (AV/VST3) for macOS. You can also build the plugin for your platform. This will require the traced models, which you can download here. First, you will need download and extract libtorch. Check the PyTorch site to find the correct version.

wget https://download.pytorch.org/libtorch/cpu/libtorch-macos-1.7.1.zip
unzip libtorch-macos-1.7.1.zip

Now move this into the realtime/ directory .

mv libtorch realtime/

We provide a ncomp.jucer file and a CMakeLists.txt that was created using FRUT. You will likely need to compile and run FRUT on this .jucer file in order to create a valid CMakeLists.txt. To do so, follow the instructions on compiling FRUT. Then convert the .jucer file. You will have to update the paths here to reflect the location of FRUT.

cd realtime/plugin/
../../FRUT/prefix/FRUT/bin/Jucer2CMake reprojucer ncomp.jucer ../../FRUT/prefix/FRUT/cmake/Reprojucer.cmake

Now you can finally build the plugin using CMake with the build.sh script. BUT, you will have to first update the path to libtorch in the build.sh script.

rm -rf build
mkdir build
cd build
cmake .. -G Xcode -DCMAKE_PREFIX_PATH=/absolute/path/to/libtorch ..
cmake --build .

Citation

If you use any of this code in your work, please consider citing us.

@inproceedings{steinmetz2022efficient,
    title={Efficient neural networks for real-time modeling of analog dynamic range compression},
    author={Steinmetz, Christian J. and Reiss, Joshua D.},
    booktitle={152nd AES Convention},
    year={2022}}