Objective Quality Assessment of Target Speaker Separation Performance in Multisource Reverberant Environment
Paco Langjahr, Pejman Mowlaee · 2013
Sound quality estimation of a speech enhancement or source separation system in a realistic adverse noise scenario is a chal-lenge. In particular, the connection between results obtained by quality metrics versus those obtained from human subjective listening tests is unknown. In this paper, as the first attempt, we present results, which examine the correlation between cer-tain objective metrics and Multi Stimulus with Hidden Refer-ence and Anchor Point (MUSHRA) listening tests. There are a variety of applications which may benefit from the outcome of this study, since the enhancement/separation system is often used as a pre-processing stage for target applications such as- to name a few- automatic speech recognition, speech cod-ing, speaker recognition and hearing aid devices. In all these applications the knowledge of which objective metrics reliably predicts subjective results is of great importance, as it saves the time consuming task of listening tests in future studies when developing a new enhancement system. The paper presents per-formance evaluation of several benchmark systems in a realis-tic noisy reverberant environment using the existing objective quality metrics. Subjective listening experiments which are rel-evant for perceptual speech quality. MUSHRA listening test is also conducted as our subjective evaluation. Comparing the ob-jective and subjective results we present the discussion about which metrics correlate well with subjective listening tests. Index Terms: Target speaker separation, objective metrics, multisource reverberant environment, subjective listening test.