Towards Pose-Invariant Audio-Visual Speech Enhancement in the Wild for Next-Generation Multi-Modal Hearing Aids

Abstract

Classical audio-visual (AV) speech enhancement (SE) and separation methods have been successful at operating under constrained environments; however, the speech quality and intelligibility improvement is significantly reduced in unconstrained real-world environments where variation in pose and illumination are encountered. In this paper, we present a novel privacy-preserving approach for real world unconstrained pose-invariant AV SE and separation that contextually exploits pose-invariant 3D landmark flow features and noisy speech features to selectively suppress unwanted background speech and non-speech noises. In addition, we present a unified architecture that integrates state-of-the-art transformers with temporal convolution neural networks for effective pose-invariant AV SE. The preliminary systematic experimentation on benchmark multi-pose OuluVS2 and LRS3-TED corpora demonstrate that the privacy preserving 3D landmark flow features are effective for pose-invariant SE and separation. In addition, the proposed AV SE model significantly outperforms state-of-the-art audio-only SE model, oracle ideal binary mask, and A-only variant of the proposed model in speaker and noise independent settings.