Overview of speech quality metrics in terms of automated evaluation of signal denoising in a presence of non-stationary noise

On Demand

Presenting Author(s)

Karol J. Duzinkiewicz

Software Engineer
Intel Technology Poland
Banino, Pomorskie, Poland

Slides Slides and Audio

Co-author(s)

PT

Pawel Trella

Intel
Gdansk, Pomorskie, Poland
DK

Damian Koszewski

Software Engineer
Gdansk, Pomorskie, Poland
Kamila Pietrusińska

Student
Gdansk University of Technology
Gdynia, Pomorskie, Poland

Abstract: Recent developments in neural network-based speech enhancement ask for a robust subjective metric that can be used for comparing performance of a different noise suppression algorithms (for non-stationary noises), that would closely match costly and time-consuming subjective user tests. The article describes results of a comparison between subjective scores obtained using MUSHRA methodology vs. automated evaluation with objective metrics i.e. POLQA, 3QUEST, STOI & ESTOI, on a set of recordings processed by 2 different denoising algorithms for close & far speaker distance. Correlation coefficient is calculated between subjective scores and examined metrics. The results are based on recordings simulated using an in-house simulation toolchain, based on impulse responses from actual laptop device used in low reverb quiet room.