Can Large Audio-Language Models Truly Hear? Tackling Hallucinations with Multi-Task Assessment and Stepwise Audio Reasoning (2410.16130v1)

Abstract: Recent advancements in large audio-LLMs (LALMs) have shown impressive capabilities in understanding and reasoning about audio and speech information. However, these models still face challenges, including hallucinating non-existent sound events, misidentifying the order of sound events, and incorrectly attributing sound sources, which undermine their reliability and real-world application. To systematically evaluate these issues, we propose three distinct tasks: object existence, temporal order, and object attribute within audio. These tasks assess the models' comprehension of critical audio information aspects. Our experimental results reveal limitations in these fundamental tasks, underscoring the need for better models in recognizing specific sound events, determining event sequences, and identifying sound sources. To improve performance in these areas, we introduce a multi-turn chain-of-thought approach, which demonstrates significantly improved model performance across the proposed tasks.

Analysis of Audio-LLM Hallucinations: A Multi-Task Approach

The paper, "Can Large Audio-LLMs Truly Hear? Tackling Hallucinations with Multi-Task Assessment and Stepwise Audio Reasoning," presents a comprehensive evaluation of large audio-LLMs (LALMs) and their propensity for hallucination. This work primarily investigates three critical areas where these models struggle: object existence, temporal order, and object attribute comprehension within audio inputs.

Challenges in Audio-LLMs

The recent evolution of LALMs offers advanced capabilities in processing audio and language data synergistically. However, these models exhibit issues such as hallucinating non-existent sounds, misinterpreting event sequences, and incorrectly attributing sound sources. Addressing these problems is imperative for the reliability of LALMs in practical applications, such as emergency response or autonomous driving.

Multi-Task Evaluation

The authors propose a framework involving three distinct tasks to systematically evaluate model hallucinations:

• Object Existence: This task examines the model's ability to detect specific sound events within audio samples.
• Temporal Order: This assessment evaluates the model's ability to identify the correct sequence of sound occurrences.
• Object Attribute: This task investigates the model's capacity to correctly attribute sounds to their respective sources.

Each task involves discriminative questioning, utilizing datasets such as AudioCaps, ESC-50, and VocalSound, among others. The authors adopt paired question sets to rigorously test the models' sensitivity to changes in audio content.

Experimental Findings

The paper reveals significant deficits in current LALMs, as reflected by their inability to consistently and accurately perform across all tasks. Performance metrics such as accuracy, precision, recall, and F1 scores show notable limitations, particularly in open-source models. The paper also highlights that these models tend to default to affirmative responses, complicating accurate audio comprehension.

MATCH Method Improvement

In response to these shortcomings, the authors introduce the Multi-turn And Thoughtful Chain of Hearings (MATCH) method. This approach first involves generating an audio description, which precedes the model's response to probing questions. The MATCH method consistently enhances performance across all tasks, substantially increasing F1 scores and improving sensitivity to temporal and attribute-related aspects of audio understanding.

Conclusion and Implications

This research underscores the urgent need for improved methodologies in LALMs to mitigate hallucination issues. The findings have profound implications for developing more reliable and accurate audio-language processing systems. The MATCH method, in particular, presents a promising avenue for enhancing the models' performance by leveraging multi-turn dialogue strategies.

Future work in this domain may focus on refining LALMs through more sophisticated audio reasoning techniques and exploring the integration of such approaches in real-world applications. The need for robust solutions remains critical as audio-LLMs continue to play pivotal roles across various sectors, from security to healthcare diagnostics.