With the rapid advancement of Multi-modal Large Language Models (MLLMs), several diagnostic benchmarks have recently been developed to assess these models' multi-modal reasoning proficiency. However, these benchmarks are restricted to assessing primarily the visual aspect and do not examine the holistic audio-visual (AV) understanding. Moreover, currently, there are no benchmarks that investigate the capabilities of AVLLMs to calibrate their responses when presented with perturbed inputs. To this end, we introduce Audio-Visual Trustworthiness assessment Benchmark (AVTrustBench), comprising 600K samples spanning over 9 meticulously crafted tasks, evaluating the capabilities of AVLLMs across three distinct dimensions: Adversarial attack, Compositional reasoning, and Modality-specific dependency. Using our benchmark we extensively evaluate 13 state-of-the-art AVLLMs. The findings reveal that the majority of existing models fall significantly short of achieving human-like comprehension, offering valuable insights for future research directions. To alleviate the limitations in the existing approaches, we further propose a robust, model-agnostic calibrated audio-visual preference optimization based training strategy CAVPref, obtaining a gain up to 30.19% across all 9 tasks. We will publicly release our code and benchmark to facilitate future research in this direction.

We present AdVerb, a novel audio-visual dereverberation framework that uses visual cues in addition to the reverberant sound to estimate clean audio. Although audioonly dereverberation is a well-studied problem, our approach incorporates the complementary visual modality to perform audio dereverberation. Given an image of the environment where the reverberated sound signal has been recorded, AdVerb employs a novel geometry-aware cross-modal transformer architecture that captures scene geometry and audio-visual cross-modal relationship to generate a complex ideal ratio mask, which, when applied to the reverberant audio predicts the clean sound. The effectiveness of our method is demonstrated through extensive quantitative and qualitative evaluations. Our approach significantly outperforms traditional audio-only and audio-visual baselines on three downstream tasks: speech enhancement, speech recognition, and speaker verification, with relative improvements in the range of 18% - 82% on the LibriSpeech test-clean set. We also achieve highly satisfactory RT60 error scores on the AVSpeech dataset.

Performing sound source separation and visual object segmentation jointly in naturally occurring videos is a notoriously difficult task, especially in the absence of annotated data. In this study, we leverage the concurrency between audio and visual modalities in an attempt to solve the joint audio-visual segmentation problem in a self-supervised manner. Human beings interact with the physical world through a few sensory systems such as vision, auditory, movement, etc. The usefulness of the interplay of such systems lies in the concept of degeneracy [1]. It tells us that the cross-modal signals can educate each other without the presence of an external supervisor. In this work, we efficiently exploit this fact that learning from one modality inherently helps to find patterns in others by introducing a novel audio-visual fusion technique. Also, to the best of our knowledge, we are the first to address the partially occluded sound source segmentation task. Our study shows that the proposed model significantly outperforms existing state-of-the-art methods in both visual and audio source separation tasks.

Electrocardiogram (ECG) signal is studied to obtain crucial information about the condition of a patient's heart. Machine learning based automated medical diagnostic systems that may help to evaluate the condition of the heart from this signal are required to be trained using large volumes of labelled training samples and the same may increase the chance of compromising with the patients' privacy. To solve this issue, generation of synthetic electrocardiogram signals by learning only from the general distributions of the available real training samples have been attempted in the literature. However, these studies did not pay necessary attention to the specific vital details of these signals, such as the P wave, the QRS complex, and the T wave. This shortcoming often results in the generation of unrealistic synthetic signals, such as a signal which does not contain one or more of the above components. In the present study, a novel deep generative architecture, termed as CardioGAN, based on generative adversarial network and powered by the effective attention mechanism has been designed which is capable of learning the intricate inter-dependencies among the various parts of real samples leading to the generation of more realistic electrocardiogram signals. Also, it helps in reducing the risk of breaching the privacy of patients. Extensive experimentation performed by us establishes that the proposed method achieves a better performance in generating synthetic electrocardiogram signals in comparison to the existing methods.