Skip to content

Latest commit

 

History

History
79 lines (56 loc) · 3.71 KB

sigsep_open-unmix-pytorch_umx.md

File metadata and controls

79 lines (56 loc) · 3.71 KB
Raw
layout background-class body-class category title summary image author tags github-link github-id featured_image_1 featured_image_2 accelerator
hub_detail
hub-background
hub
researchers
Open-Unmix
Reference implementation for music source separation
sigsep_logo_inria.png
Inria
audio
scriptable
sigsep/open-unmix-pytorch
sigsep_umx-diagram.png
no-image
cuda-optional
 
# assuming you have a PyTorch >=1.6.0 installed
pip install -q torchaudio
import torch

# loading umxhq four target separator
separator = torch.hub.load('sigsep/open-unmix-pytorch', 'umxhq')

# generate random audio
# ... with shape (nb_samples, nb_channels, nb_timesteps)
# ... and with the same sample rate as that of the separator
audio = torch.rand((1, 2, 100000))
original_sample_rate = separator.sample_rate

# make sure to resample the audio to models' sample rate, separator.sample_rate, if the two are different
# resampler = torchaudio.transforms.Resample(original_sample_rate, separator.sample_rate)
# audio = resampler(audio)

estimates = separator(audio)
# estimates.shape = (1, 4, 2, 100000)

Model Description

Open-Unmix provides ready-to-use models that allow users to separate pop music into four stems: vocals, drums, bass and the remaining other instruments. The models were pre-trained on the freely available MUSDB18 dataset.

Each target model is based on a three-layer bidirectional deep LSTM. The model learns to predict the magnitude spectrogram of a target source, like vocals, from the magnitude spectrogram of a mixture input. Internally, the prediction is obtained by applying a mask on the input. The model is optimized in the magnitude domain using mean squared error.

A Separator meta-model (as shown in the code example above) puts together multiple Open-unmix spectrogram models for each desired target, and combines their output through a multichannel generalized Wiener filter, before application of inverse STFTs using torchaudio. The filtering is differentiable (but parameter-free) version of norbert.

Pre-trained Separator models

  • umxhq (default) trained on MUSDB18-HQ which comprises the same tracks as in MUSDB18 but un-compressed which yield in a full bandwidth of 22050 Hz.

  • umx is trained on the regular MUSDB18 which is bandwidth limited to 16 kHz do to AAC compression. This model should be used for comparison with other (older) methods for evaluation in SiSEC18.

Furthermore, we provide a model for speech enhancement trained by Sony Corporation

All three models are also available as spectrogram (core) models, which take magnitude spectrogram inputs and ouput separated spectrograms. These models can be loaded using umxhq_spec, umx_spec and umxse_spec.

Details

For additional examples, documentation and usage examples, please visit this the github repo.

Furthermore, the models and all utility function to preprocess, read and save audio stems, are available in a python package that can be intalled via

pip install openunmix

References