fast_bss_eval

Do you have a zillion BSS audio files to process and it is taking days ? Is your simulation never ending ?

Fear no more! fast_bss_eval is here to help you!

fast_bss_eval is a fast implementation of the bss_eval metrics for the evaluation of blind source separation. Our implementation of the bss_eval metrics has the following advantages compared to other existing ones.

• seamlessly works with both numpy arrays and pytorch tensors
• very fast
• can be even faster by using an iterative solver (add use_cg_iter=10 option to the function call)
• differentiable via pytorch
• can run on GPU via pytorch

Author

• Robin Scheibler

Quick Start

Install

# from pypi
pip install fast-bss-eval

# or from source
git clone https://github.com/fakufaku/fast_bss_eval
cd fast_bss_eval
pip install -e .

Use

Assuming you have multichannel signals for the estmated and reference sources stored in wav format files names my_estimate_file.wav and my_reference_file.wav, respectively, you can quickly evaluate the bss_eval metrics as follows.

from scipy.io import wavfile
import fast_bss_eval

# open the files, we assume the sampling rate is known
# to be the same
fs, ref = wavfile.read("my_reference_file.wav")
_, est = wavfile.read("my_estimate_file.wav")

# compute the metrics
sdr, sir, sar, perm = fast_bss_eval.bss_eval_sources(ref.T, est.T)

Benchmark

This package is significantly faster than other packages that also allow to compute bss_eval metrics such as mir_eval or sigsep/bsseval. We did a benchmark using numpy/torch, single/double precision floating point arithmetic (fp32/fp64), and using either Gaussian elimination or a conjugate gradient descent (solve/CGD10).

License

2021 (c) Robin Scheibler, LINE Corporation

This code is released under MIT License.