ci-sdr

Convolutive Transfer Function Invariant SDR

This repository contains an implementation for the Convolutive transfer function Invariant Signal-to-Distortion Ratio objective for PyTorch as described in the publication Convolutive Transfer Function Invariant SDR training criteria for Multi-Channel Reverberant Speech Separation (link arXiv).

Here, a small example, how you can use this CI-SDR objective in your own source code:

import torch
import ci_sdr

reference: torch.tensor = ...
# reference.shape: [speakers, samples]

estimation: torch.tensor = ...
# estimation shape: [speakers, samples]

sdr = ci_sdr.pt.ci_sdr_loss(estimation, reference)
# sdr shape: [speakers]

The idea of this objective function is based in the theory from E. Vincent, R. Gribonval and C. Févotte, Performance measurement in blind audio source separation, IEEE Trans. Audio, Speech and Language Processing, known as BSSEval. The original author provided MATLAB source code (link) and the package mir_eval (link) contains a python port. Some peoble refer to these implementations as BSSEval v3 (link).

The PyTorch code in this package is tested to yield the same SDR values as mir_eval with the default parameters.

NOTE: If you want to use BSSEval v3 SDR as metric, I recomment to use mir_eval.separation.bss_eval_sources and use as reference the clean/unreverberated source signals. The implementation in this repository has minor difference that makes it problematic to compare SDR values accorss different publications (e.g. here the permutation is calculated on the SDR, while mir_eval computes it based on the SIR.).

Installation

Install it directly with Pip, if you just want to use it:

pip install ci-sdr

or to get the recent version:

pip install git+https://github.com/fgnt/ci_sdr.git

If you want to install it with all dependencies (test and doctest dependencies), run:

pip install git+https://github.com/fgnt/ci_sdr.git#egg=ci_sdr[all]

When you want to change the code, clone this repository and install it as editable:

git clone https://github.com/fgnt/ci_sdr.git
cd ci_sdr
pip install --editable .
# pip install --editable .[all]

Citation

To cite this implementation, you can cite the following paper (link):

@article{boeddeker2020convolutive,
  title   = {Convolutive Transfer Function Invariant {SDR} training criteria for Multi-Channel Reverberant Speech Separation},
  author  = {Boeddeker, Christoph and Zhang, Wangyou and Nakatani, Tomohiro and Kinoshita, Keisuke and Ochiai, Tsubasa and Delcroix, Marc and Kamo, Naoyuki and Qian, Yanmin and Haeb-Umbach, Reinhold},
  journal = {arXiv preprint arXiv:2011.15003},
  year    = {2020}
}