Evaluation of Objective Quality Models on Neural Audio Codecs

With the emergence of neural audio codecs and new objective quality models based on machine learning, there is a need to clarify which models predict accurately the perceptual quality of coded speech. In this paper, we consider a selected sub-set of ten objective quality models; we present a correlation analysis based on test results from a P.800 ACR experiment on clean speech, assessing the quality of neural speech/audio codecs – traditional codecs (EVS, Opus) are also included as yardsticks. The evaluation is limited to signal-based models for listening-only quality. Overall results per condition are analyzed in terms of Pearson’s correlation, Kendall’s Tau and root mean squared error (RMSE); objective scores per codec are also discussed.