Statistical selection of relevant objective criteria for speech enhancement assessment

The quality of denoised speeches is ideally evaluated using subjective listening tests. But, to facilitate the heavy workload, objective criteria represent an efficient alternative. Hence, many criteria are developed in the literature and more than one criterion is usually used. In this paper, we propose to address the problem of classical objective criteria choice to select those which are relevant for evaluating subjective quality of denoised signals. Two tools are used: the boxplots to discard criteria leading to a confusion when trying to discriminate between the five classes of MOS criteria and the Principal Component Analysis (PCA) to eliminate redundancy between criteria and to reduce dimensionality.