See hear now: is audio-visual QoE now just a fusion of audio and video metrics?

Single-modal audio/speech and video quality models have reached high levels of performance. Although traditional algorithms are still preferred for many practical applications, advances in machine learning (ML) and deep learning techniques have exceeded their performance in several scientific comparisons. However, audio-visual (AV) models have received signifi-cantly less attention and development. Despite the acknowledged challenge that multimodal interaction poses to the AV problem, traditional AV models generally rely on simple fusion techniques of individual audio and video predictions. Consequently, the impact of recent advances in single-modal quality assessment models on SOTA (state-of-the-art) AV quality models merits attention. This paper presents a revised and updated benchmark for AV quality assessment with particular focus on new speech quality metrics. Three AV datasets were used to test audio, video, and AV quality metrics. For audio and video, the best performing metrics were selected to build simple late-fusion models using their raw predictions. The fused models were then compared to the SOTA AV models. Results show that a simple fusion strategy produces accurate AV quality predictions (LCC and SCC greater than 0.90) with low error rates (RMSE lower than 0.33). These results highlight the influence of advances in speech quality for AV quality assessment.