Includes reverberation, distortion, dynamic range processing, equalization, stereo processing.
Enables virtual analog modeling, blind parameter estimation, automated DSP, and style transfer.
Batching with operation on both CPU and GPU accelerators for fast training and reduced bottlenecks.
Open source and free to use for academic and commercial applications under Apache 2.0 license.
pip install dasp-pytorch
Or, for a local installation.
git clone https://github.com/csteinmetz1/dasp-pytorch
cd dasp-pytorch
pip install -e .
dasp-pytorch is a Python library for constructing differentiable audio signal processors using PyTorch.
These differentiable processors can be used standalone or within the computation graph of neural networks.
We provide purely functional interfaces for all processors that enable ease-of-use and portability across projects.
Unless oterhwise stated, all effect functions expect 3-dim tensors with shape (batch_size, num_channels, num_samples) as input and output.
Using an effect in your computation graph is as simple as calling the function with the input tensor as argument.
Here is a minimal example to demonstrate reverse engineering the drive value of a simple distortion effect using gradient descent.
Try it for yourself:
import torch
import torchaudio
import dasp_pytorch
# Load audio
x, sr = torchaudio.load("audio/short_riff.wav")
# create batch dim
# (batch_size, n_channels, n_samples)
x = x.unsqueeze(0)
# apply some distortion with 16 dB drive
drive = torch.tensor([16.0])
y = dasp_pytorch.functional.distortion(x, sr, drive)
# create a parameter to optimizer
drive_hat = torch.nn.Parameter(torch.tensor(0.0))
optimizer = torch.optim.Adam([drive_hat], lr=0.01)
# optimize the parameter
n_iters = 2500
for n in range(n_iters):
# apply distortion with the estimated parameter
y_hat = dasp_pytorch.functional.distortion(x, sr, drive_hat)
# compute distance between estimate and target
loss = torch.nn.functional.mse_loss(y_hat, y)
# optimize
optimizer.zero_grad()
loss.backward()
optimizer.step()
print(
f"step: {n+1}/{n_iters}, loss: {loss.item():.3e}, drive: {drive_hat.item():.3f}\r"
)
For the remaining examples we will use the GuitarSet dataset. You can download the data using the following commands:
mkdir data
wget https://zenodo.org/records/3371780/files/audio_mono-mic.zip
unzip audio_mono-mic.zip
rm audio_mono-mic.zip| Audio Processor | Functional Interface |
|---|---|
| Gain | gain() |
| Distortion | distortion() |
| Parametric Equalizer | parametric_eq() |
| Dynamic range compressor | compressor() |
| Dynamic range expander | expander() |
| Reverberation | noise_shaped_reverberation() |
| Stereo Widener | stereo_widener() |
| Stereo Panner | stereo_panner() |
| Stereo Bus | stereo_bus() |
If you use this library consider citing these papers:
Differentiable parametric EQ and dynamic range compressor
@article{steinmetz2022style,
title={Style transfer of audio effects with differentiable signal processing},
author={Steinmetz, Christian J and Bryan, Nicholas J and Reiss, Joshua D},
journal={arXiv preprint arXiv:2207.08759},
year={2022}
}Differentiable artificial reveberation with frequency-band noise shaping
@inproceedings{steinmetz2021filtered,
title={Filtered noise shaping for time domain room impulse
response estimation from reverberant speech},
author={Steinmetz, Christian J and Ithapu, Vamsi Krishna and Calamia, Paul},
booktitle={WASPAA},
year={2021},
organization={IEEE}
}Differentiable IIR filters
@inproceedings{nercessian2020neural,
title={Neural parametric equalizer matching using differentiable biquads},
author={Nercessian, Shahan},
booktitle={DAFx},
year={2020}
}@inproceedings{colonel2022direct,
title={Direct design of biquad filter cascades with deep learning
by sampling random polynomials},
author={Colonel, Joseph T and Steinmetz, Christian J and
Michelen, Marcus and Reiss, Joshua D},
booktitle={ICASSP},
year={2022},
organization={IEEE}Supported by the EPSRC UKRI Centre for Doctoral Training in Artificial Intelligence and Music (EP/S022694/1).
