A multitask teacher-student framework for perceptual audio quality assessment

Introduction

This is the complementary repository for the following publication:

Chih-Wei Wu, Phillip A. Williams, William Wolcott, 
"A multitask teacher-student framework for perceptual audio quality assessment",
in Proceedings of European Signal Processing Conference (EUSIPCO), 2021

Specifically, this repository includes:

1. Additional metrics (Please see the sections below)

2. Additional plots/figures (Please see the sections below)

3. Software implementation and pretrained model for the proposed system (Please see here for more details.)

Additional Metrics

Root mean square error (RMSE) was computed as a secondrary metric but not shown in the manuscript due to the space constraints.

Systems			Test Datasets			Averaged RMSE
	BS1387Conform	CoreSv14	UnbAvq2013	NLow	NHigh
HAAQI	0.913	0.886	0.717	1.601	0.278	0.879
PEAQ	0.484	0.440	1.183	1.669	0.484	0.852
PEMO-Q	0.608	0.638	1.147	1.292	0.242	0.786
ViSQOLAudio	0.697	0.982	0.597	0.492	0.497	0.653
STL-HAAQI	0.927	0.858	0.645	1.440	0.298	0.834
STL-PEAQ	1.125	0.791	1.041	0.966	0.546	0.894
STL-PEMO-Q	0.600	0.577	1.208	0.926	0.254	0.713
STL-ViSQOLAudio	0.714	0.992	0.618	0.528	0.596	0.690
MTL-T1	0.818	1.009	0.689	1.408	0.350	0.855
MTL-T2	0.960	0.871	1.082	1.024	0.687	0.925
MTL-T3	0.503	0.575	1.311	0.920	0.200	0.702
MTL-T4	0.604	0.943	0.565	0.496	0.463	0.614
MTL-mean	0.582	0.746	0.758	0.805	0.281	0.634
MTL-gmean	0.524	0.713	0.904	0.645	0.295	0.616
MTL-median	0.577	0.758	0.721	0.785	0.284	0.625

Additional Plots/Figures

The following scatter plots show the model predictions v.s. human ground truth on the combined proprietary (i.e., NLow + NHigh) datasets. Here we compare the best performing teacher system (ViSQOLAudio) and the best performing student system (MTL-gmean):

Fig. 1. ViSQOLAudio on {NLow + NHigh}

Fig. 2. MTL-gmean on {NLow + NHigh}

Citing

To cite this work, please use the following:

@inproceedings{
  wu_2021,
  title={A multitask teacher-student framework for perceptual audio quality assessment},
  author={Wu, Chih-Wei and Williams, Phillip A and Wolcott, William},
  booktitle={European Signal Processing Conference (EUSIPCO)},
  year={2021}
}

Contact

Chih-Wei Wu
cwu307[at]gmail.com