Towards Entertainment Live-stream Sound Quality Assessment: A High-Quality Benchmark and Non-Intrusive Multi-Order Contextual Attention Framework

To date, many sound quality assessment algorithms have been proposed to address speech quality prediction problem. However, utilizing these methods in real-world entertainment live-stream scenarios directly may suffer from the following limitations: (1) There are no publicly available high-quality annotated sound datasets gathered from real-world entertainment live-stream platform to train and verify the effectiveness of the model. (2) The existing sound quality assessment methods are not designed for entertainment live-stream scenarios. To overcome these limitations, in this work, we first present a High-quality Entertainment Live-stream Sound (HEL-Sound) dataset, which contains 18092 sound files with high-quality annotated Mean Opinion Score (MOS), gathered from the real-world entertainment live-stream platform. To the best of our knowledge, it is the largest real-world entertainment live-stream sound quality assessment dataset to date. Meanwhile, we design the Non-intrusive Multi-order Contextual Attention (NiMCA) framework to construct a strong baseline, which jointly modeling the cross-order context information and the complex semantics of entertainment live-stream sound in a unified deep model under non-intrusive setting. Extensive experiments conducted on the collected HEL-Sound demonstrate that our proposed NiMCA achieves state-of-the-art performance.