We provide a PyTorch implementation of the paper: Real Time Speech Enhancement in the Waveform Domain. In which, we present a causal speech enhancement model working on the raw waveform that runs in real-time on a laptop CPU. The proposed model is based on an encoder-decoder architecture with skip-connections. It is optimized on both time and frequency domains, using multiple loss functions. Empirical evidence shows that it is capable of removing various kinds of background noise including stationary and non-stationary noises, as well as room reverb. Additionally, we suggest a set of data augmentation techniques applied directly on the raw waveform which further improve model performance and its generalization abilities.
Audio samples can be found here: Samples
The proposed model is based on the Demucs architecture, originally proposed for music source-separation: (Paper, Code).
First, install Python 3.7 (recommended with Anaconda).
Just run
pip install denoiserClone this repository and install the dependencies. We recommend using a fresh virtualenv or Conda environment.
git clone https://github.com/facebookresearch/denoiser
cd denoiser
pip install -r requirements.txt # If you don't have cuda
pip install -r requirements_cuda.txt # If you have cudaIf you want to use denoiser live (for a Skype call for instance), you will
need a specific loopback audio interface.
On Mac OS X, this is provided by Soundflower. First install Soundflower, and then you can just run
python -m denoiser.liveIn your favorite video conference call application, just select "Soundflower (2ch)" as input to enjoy your denoised speech.
Watch our live demo presentation in the following link: Demo.
At the moment, we do not provide official support for other OSes. However, if you
have a a soundcard that supports loopback (for instance Steinberg products), you can try
to make it work. You can list the available audio interfaces with python -m sounddevice.
Then once you have spotted your loopback interface, just run
python -m denoiser.live --out INDEX_OR_NAME_OF_LOOPBACK_IFACEBy default, denoiser will use the default audio input. You can change that with the --in flag.
Note that on Windows you will need to replace python by python.exe.
denoiser can introduce distortions for very high level of noises.
Audio can become crunchy if your computer is not fast enough to process audio in real time.
In that case, you will see an error message in your terminal warning you that denoiser
is not processing audio fast enough. You can try exiting all non required applications.
denoiser was tested on a Mac Book Pro with an 2GHz quadcore Intel i5 with DDR4 memory.
You might experience issues with DDR3 memory. In that case you can trade overall latency for speed by processing multiple frames at once. To do so, run
python -m denoiser.live -f 2
You can increase to -f 3 or more if needed, but each increase will add 16ms of extra latency.
You can also denoise received speech, but you won't be able to both denoise your own speech and the received speech (unless you have a really beefy computer and enough loopback audio interfaces). This can be achieved by selecting the loopback interface as the audio output of your VC software and then running
python -m denoiser.live --in "Soundflower (2ch)" --out "NAME OF OUT IFACE"- Run
sh make_debug.shto generate json files for the toy dataset. - Run
python train.py
We use Hydra to control all the training configurations. If you are not familiar with Hydra we recommend visiting the Hydra website. Generally, Hydra is an open-source framework that simplifies the development of research applications by providing the ability to create a hierarchical configuration dynamically.
The config file with all relevant arguments for training our model can be found under the conf folder.
Notice, under the conf folder, the dset folder contains the configuration files for
the different datasets. You should see a file named debug.yaml with the relevant configuration for the debug sample set.
You can pass options through the
command line, for instance ./train.py demucs.hidden=32.
Please refer to conf/config.yaml for a reference of the possible options.
You can also directly edit the config.yaml file, although this is not recommended
due to the way experiments are automatically named, as explained hereafter.
Each experiment will get a unique name based on the command line options you passed.
Restarting the same command will reuse the existing folder and automatically
start from a previous checkpoint if possible. In order to ignore previous checkpoints,
you must pass the restart=1 option.
Note that options like device, num_workers, etc. have no influence on the experiment name.
If you want to train using a new dataset, you can:
- Create a separate config file for it.
- Place the new config files under the
dsetfolder. Check conf/dset/debug.yaml for more details on configuring your dataset. - Point to it either in the general config file or via the command line, e.g.
./train.py dset=name_of_dset.
You also need to generate the relevant .jsonfiles in the egs/folder.
For that purpose you can use the python -m denoiser.audio command that will
scan the given folders and output the required metadata as json.
For instance, if your noisy files are located in $noisy and the clean files in $clean, you can do
out=egs/mydataset/tr
mkdir -p $out
python -m denoiser.audio $noisy > $out/noisy.json
python -m denoiser.audio $clean > $out/clean.jsonThe data loader reads both clean and noisy json files named: clean.json and noisy.json. These files should contain all the paths to the wav files to be used to optimize and test the model along with their size (in frames).
You can use python -m denoiser.audio FOLDER_WITH_WAV1 [FOLDER_WITH_WAV2 ...] > OUTPUT.json to generate those files.
You should generate the above files for both training and test sets (and validation set if provided). Once this is done, you should create a yaml (similarly to conf/dset/debug.yaml) with the dataset folders' updated paths.
Please check conf/dset/debug.yaml for more details.
Training is simply done by launching the train.py script:
./train.py
This scripts read all the configurations from the conf/config.yaml file.
To launch distributed training you should turn on the distributed training flag. This can be done as follows:
./train.py ddp=1
Logs are stored by default in the outputs folder. Look for the matching experiment name.
In the experiment folder you will find the best.th serialized model, the training checkpoint checkpoint.th,
and well as the log with the metrics trainer.log. All metrics are also extracted to the history.json
file for easier parsing. Enhancements samples are stored in the samples folder (if noisy_dir or noisy_json
is set in the dataset).
You can fine-tune one of the 3 pre-trained models dns48, dns64 and master64. To do so:
./train.py continue_pretrained=dns48
./train.py continue_pretrained=dns64 demucs.hidden=64
./train.py continue_pretrained=master64 demucs.hidden=64
Evaluating the models can be done by:
python -m denoiser.evaluate --model_path=<path to the model> --data_dir=<path to folder containing noisy.json and clean.json>
Note that the path given to --model_path should be obtained from one of the best.th file, not checkpoint.th.
It is also possible to use pre-trained model, using either --dns48, --dns64or --master64.
For more details regarding possible arguments, please see:
usage: denoiser.evaluate [-h] [-m MODEL_PATH | --dns48 | --dns64 | --master64]
[--device DEVICE] [--dry DRY]
[--sample_rate SAMPLE_RATE]
[--num_workers NUM_WORKERS] [--streaming]
[--data_dir DATA_DIR] [--matching MATCHING]
[--no_pesq] [-v]
Speech enhancement using Demucs - Evaluate model performance
optional arguments:
-h, --help show this help message and exit
-m MODEL_PATH, --model_path MODEL_PATH
Path to local trained model.
--dns48 Use pre-trained real time H=48 model trained on DNS.
--dns64 Use pre-trained real time H=64 model trained on DNS.
--master64 Use pre-trained real time H=64 model trained on DNS
and Valentini.
--device DEVICE
--dry DRY dry/wet knob coefficient. 0 is only input signal, 1
only denoised.
--sample_rate SAMPLE_RATE
sample rate
--num_workers NUM_WORKERS
--streaming true streaming evaluation for Demucs
--data_dir DATA_DIR directory including noisy.json and clean.json files
--matching MATCHING set this to dns for the dns dataset.
--no_pesq Don't compute PESQ.
-v, --verbose More loggging
Generating the enhanced files can be done by:
python -m denoiser.enhance --model_path=<path to the model> --noisy_dir=<path to the dir with the noisy files> --out_dir=<path to store enhanced files>
Notice, you can either provide noisy_dir or noisy_json for the test data.
Note that the path given to --model_path should be obtained from one of the best.th file, not checkpoint.th.
It is also possible to use pre-trained model, using either --dns48, --dns64or --master64.
For more details regarding possible arguments, please see:
usage: denoiser.enhance [-h] [-m MODEL_PATH | --dns48 | --dns64 | --master64]
[--device DEVICE] [--dry DRY]
[--sample_rate SAMPLE_RATE]
[--num_workers NUM_WORKERS] [--streaming]
[--out_dir OUT_DIR] [--batch_size BATCH_SIZE] [-v]
[--noisy_dir NOISY_DIR | --noisy_json NOISY_JSON]
Speech enhancement using Demucs - Generate enhanced files
optional arguments:
-h, --help show this help message and exit
-m MODEL_PATH, --model_path MODEL_PATH
Path to local trained model.
--dns48 Use pre-trained real time H=48 model trained on DNS.
--dns64 Use pre-trained real time H=64 model trained on DNS.
--master64 Use pre-trained real time H=64 model trained on DNS
and Valentini.
--device DEVICE
--dry DRY dry/wet knob coefficient. 0 is only input signal, 1
only denoised.
--sample_rate SAMPLE_RATE
sample rate
--num_workers NUM_WORKERS
--streaming true streaming evaluation for Demucs
--out_dir OUT_DIR directory putting enhanced wav files
--batch_size BATCH_SIZE
batch size
-v, --verbose more loggging
--noisy_dir NOISY_DIR
directory including noisy wav files
--noisy_json NOISY_JSON
json file including noisy wav files
Here we provide a detailed description of how to reproduce the results from the paper:
- Download Valentini dataset.
- Adapt the Valentini config file and run the processing script.
- Generate the egs/ files as explained here after.
- Launch the training using the
launch_valentini.shscript.
To create the egs/ file, adapt and run the following code
noisy_train=path to valentini
clean_train=path to valentini
noisy_test=path to valentini
clean_test=path to valentini
noisy_dev=path to valentini
clean_dev=path to valentini
mkdir -p egs/val/tr
mkdir -p egs/val/cv
mkdir -p egs/val/tt
python -m denoiser.audio $noisy_train egs/val/tr/noisy.json
python -m denoiser.audio $clean_train egs/val/tr/clean.json
python -m denoiser.audio $noisy_test egs/val/tt/noisy.json
python -m denoiser.audio $clean_test egs/val/tt/clean.json
python -m denoiser.audio $noisy_dev egs/val/cv/noisy.json
python -m denoiser.audio $clean_dev egs/val/cv/clean.json- Download both DNS dataset.
- Adapt the DNS config file and run the processing script.
- Generate the egs/ files as explained here after.
- Launch the training using the
launch_dns.shscript.
To create the egs/ file, adapt and run the following code
dns=path to dns
noisy=path to processed noisy
clean=path to processed clean
testset=$dns/datasets/test_set
mkdir -p egs/dns/tr
python -m denoiser.audio $noisy egs/dns/tr/noisy.json
python -m denoiser.audio $clean egs/dns/tr/clean.json
mkdir -p egs/dns/tt
python -m denoiser.audio $testset/synthetic/no_reverb/noisy $testset/synthetic/reverb/noisy > egs/dns/tt/noisy.json
python -m denoiser.audio $testset/synthetic/no_reverb/clean $testset/synthetic/reverb/clean > egs/dns/tt/clean.jsonOur online implementation is based on pure python code with some optimization of the streaming convolutions and transposed convolutions. We benchmark this implementation on a quad-core Intel i5 CPU at 2 GHz. The Real-Time Factor (RTF) of the proposed models are:
| Model | Threads | RTF |
|---|---|---|
| H=48 | 1 | 0.8 |
| H=64 | 1 | 1.2 |
| H=48 | 4 | 0.6 |
| H=64 | 4 | 1.0 |
In order to compute the RTF on your own CPU launch the following command:
python -m denoiser.demucs --hidden=48 --num_threads=1
The output should be something like this:
total lag: 41.3ms, stride: 16.0ms, time per frame: 12.2ms, delta: 0.21%, RTF: 0.8
Feel free to explore different settings, i.e. bigger models and more CPU-cores.
If you use the code in your paper, then please cite it as:
@inproceedings{defossez2020real,
title={Real Time Speech Enhancement in the Waveform Domain},
author={Defossez, Alexandre and Synnaeve, Gabriel and Adi, Yossi},
booktitle={Interspeech},
year={2020}
}
This repository is released under the CC-BY-NC 4.0. license as found in the LICENSE file.
The file denoiser/stft_loss.py was adapted from the kan-bayashi/ParallelWaveGAN repository. It is an unofficial implementation of the ParallelWaveGAN paper, released under the MIT License.
The file scripts/matlab_eval.py was adapted from the santi-pdp/segan_pytorch repository. It is an unofficial implementation of the SEGAN paper, released under the MIT License.