SEP_EVAL: a collection of speech enhancement measures

This small library collects some measures for source separation and speech enhancement. The library contains - SDR - STOI - Extended STOI - PESQ (see below for details on installation) - BSS_Eval measures

Most of them are already present as python packages. This library only collects them and allows to compute all of them in one single step even on batched data.

Installation

Simply

pip install sep_eval

or directly from here

git clone https://github.com/Enny1991/sep_eval
cd sep_eval
python setup.py install

PESQ

Unfortunately, to get pesq scores, there is not other way than using the binary files from ITU. Follow the next steps:

Download the zip file from here

unzip T-REC-P.862-200511-I\!Amd2\!SOFT-ZST-E
cd Software
unzip P862_annex_A_2005_CD\ \ wav\ final.zip 
cd P862_annex_A_2005_CD/source/
gcc -o PESQ *.c -lm

The compilation will produce a unique file called pesq in this folder. Add this location to $PATH

Usage

Simply call the desired measure (or full_eval) with degraded and reference signals. Degraded and reference can also be arrays or lists and you can choose to get a full result or an average across all the given samples (see example1 and example2 for more details)

import sep_eval
import soundfile as sf

degraded, fs = sf.read('../wavs/degraded.wav')
reference, _ = sf.read('../wavs/reference1.wav')


eval_deg = sep_eval.full_eval(degraded, reference)  # returns a dictionary will all the measures

pesq_deg = sep_eval.pesq(degraded, reference)

stoi_deg = sep_eval.stoi(degraded, reference)