A Robust, Explainable, and Efficient Deep Metric for Speech and General Audio Quality

We present a robust Speech and Audio Quality Metric (SAQM) that merges features of traditional quality metrics with a CNN trained on a large amount of data using contrastive learning methods. To overcome the problem of limited amounts of human-labeled training data, we propose two pretraining strategies relying on pseudo-labels. One uses an established objective metric PEAQ-2f (based on ITU-R BS.1387) as surrogate loss, the other is based on contrasting a wide range of coding modes by a rule-of-thumb ordering based on expected perceived quality. As a key novelty, our approach lets us generate saliency maps by propagating gradients to identify which feature has the greatest impact on the final metric outcome and enables our system to provide insights regarding the type and extent of distortions present in the tested audio signal. The resulting (SAQM) operates at low computational complexity, performs comparably to (SOTA) metrics on speech while still performing well on general audio content.