About the nussl External File Zoo

This website houses and distributes extra files that are too large to ship with nussl, such as trained deep learning models, audio files, and benchmark files for testing. Please see the contributions section on Github to contribute files associated with your algorithms.

Usage

Through nussl API

Recommended

The best and easiest way to download these files is through nussl itself. The download functions in nussl will download the desired file from this website (if the file is not already on your computer) and when finished the functions will return the full path to the file.

Click here to see the official documentation for how to use the nussl download utilities.

Example 1

You can see what files are available through the helper functions in nussl.utils. For instance, to see what audio files are available simply do:

>>> import nussl
>>> nussl.utils.print_available_audio_files()
File Name                                Duration (sec)  Size       Description
dev1_female3_inst_mix.wav                10.0            1.7MiB     Instantaneous mixture of three female speakers talking in a stereo field.
dev1_female3_synthconv_130ms_5cm_mix.wav 10.0            1.7MiB     Three female speakers talking in a stereo field, with 130ms of inter-channel delay.
K0140.wav                                5.0             431.0KiB   Acoustic piano playing middle C.
K0149.wav                                5.0             430.0KiB   Acoustic piano playing the A above middle C. (A440)

This message has the same information that is shown on this website.

To download one of these files, just give a file name to the associated download function like so:

>>> audio_path = nussl.utils.download_audio_file('K0140.wav')
Saving file at ~/.nussl/audio/K0140.wav
Downloading K0140.wav from https://nussl.ci.northwestern.edu/static/audio/K0140.wav
K0140.wav...100%
>>> print(audio_path)
~/.nussl/audio/K0140.wav

The download function returns a path to the audio file.

Example 2

There are analogous for downloading trained models and benchmarks. These functions work the same as the audio example above; you can see what files are available and descriptions of them, as well as download them. For example, to download a model do the following:

>>> model_path = nussl.utils.download_trained_model('example.model')
Saving file at ~/.nussl/models/example.model
Downloading example.model from https://nussl.ci.northwestern.edu/static/trained_models/example.model
example.model...100%
>>> print(model_path)
~/.nussl/models/example.model

External File Zoo

Filename Description Size
Download
K0140.wav Acoustic piano playing middle C. 431.0KiB
K0149.wav Acoustic piano playing A above middle C (A440). 430.0KiB
dev1_female3_inst_mix.wav Instantaneous mixture of three female speakers talking in a stereo field. 1.7MiB
dev1_female3_synthconv_130ms_5cm_mix.wav Three female speakers talking in a stereo field, with 130ms of inter-channel delay. 1.7MiB
mix_3s.wav Drums and flute. 258.4KiB
drums.wav Drums playing a simple beat. 861.4KiB
flute.wav Flute playing a melody repeated twice (with slightly altered articulation). 947.5KiB
mix1.wav Mixture of flute, drums, and female speaker. 861.4KiB
mix2.wav Mixture of flute, drums, and female speaker in stereo. 1.7MiB
mix3.wav Mixture of flute and drums. 1.7MiB
mix4.wav Mixture of flute and drums. Alt. 861.4KiB
dev1_wdrums_inst_mix.wav Instantaneous mixture of flute and drums in a stereo field. 625.0KiB
wsj_speech_mixture_ViCfBJj.mp3 Two speaker mixture from WSJ0-2mix. 42.2KiB
schoolboy_fascination_excerpt.wav Excerpt from MUSDB - Al James - Schoolboy Fascination 2.5MiB
Filename Description Size
Download
speech_wsj8k.pth Deep clustering model trained on 2 speaker mixtures from WSJ0 at a sampling rate of 8000 Hz. 34.6MiB
untrained.pth Deep clustering model barely trained (1 epoch) on 2 speaker mixtures from WSJ0 at a sampling rate of 8000 Hz 307.2KiB
vocals_44k.pth Deep clustering model trained for music/voice separation on MUSDB. 55.8MiB
Filename Description Size
Download
mix3_matlab_repet_background_bRuDiWq.mat Background matrix for Repet class benchmark test. 6.5MiB
mix3_matlab_repet_foreground.mat Foreground matrix for Repet class benchmark test. 6.4MiB
benchmark_atn_bins.npy Attenuation bins numpy array for DUET benchmark test. 488.0B
benchmark_atn_delay_est.npy Attenuation/Delay estimation numpy array for DUET benchmark test. 128.0B
benchmark_atn_peak.npy Attenuation peak numpy array for DUET benchmark test. 104.0B
benchmark_delay.npy Delay numpy array for DUET benchmark test. 3.4MiB
benchmark_delay_bins.npy Delay bins for DUET benchmark tests. 488.0B
benchmark_delay_peak.npy Delay histogram peaks for DUET benchmarks 104.0B
benchmark_hist.npy Complete attenuation/delay histogram for the DUET benchmark test. 19.6KiB
benchmark_masks.npy Result masks for the DUET benchmark test. 1.3MiB
benchmark_peak_indices.npy Selected peaks on the AD histogram for DUET benchmark test. 128.0B
benchmark_stft_ch0.npy STFT channel 0 for the DUET benchmark test. 6.8MiB
benchmark_stft_ch1.npy STFT channel 1 for the DUET benchmark test. 6.8MiB
benchmark_sym_atn.npy Symmetric attenuation histogram for the DUET benchmark test. 3.4MiB
benchmark_wmat.npy Frequency matrix for the DUET benchmark test. 3.4MiB