HalluAudio: A Comprehensive Benchmark for Hallucination Detection in Large Audio-Language Models
Large Audio-Language Models (LALMs) have recently made significant strides in performance across various audio-centric tasks. However, a critical challenge that persists in this field is the issue of hallucination. Hallucination occurs when models generate responses that are semantically incorrect or lack acoustic support. This phenomenon has been largely underexplored in the audio domain, creating a gap in the understanding of LALMs’ capabilities.
Introduction to HalluAudio
The existing benchmarks for hallucination primarily focus on text or vision, and the few studies oriented towards audio are limited in scale, modality coverage, and diagnostic depth. To address this gap, the research community has introduced HalluAudio, the first large-scale benchmark designed explicitly for evaluating hallucinations across various audio modalities, including speech, environmental sound, and music.
Key Features of HalluAudio
HalluAudio comprises over 5,000 human-verified question-and-answer pairs and covers a diverse array of task types. The benchmark is structured to systematically induce hallucinations through innovative methodologies. Here are some of the standout features:
- Diverse Task Types: HalluAudio includes binary judgments, multi-choice reasoning, attribute verification, and open-ended question-answering tasks.
- Systematic Induction of Hallucinations: The research team designed adversarial prompts and mixed-audio conditions to effectively elicit hallucinations in LALMs.
- Comprehensive Evaluation Protocol: Beyond assessing accuracy, the evaluation framework measures hallucination rate, yes/no bias, error-type analysis, and refusal rate, allowing for a nuanced understanding of LALM failure modes.
Benchmarking Results
The introduction of HalluAudio also enables the benchmarking of a broad range of open-source and proprietary models, providing an unprecedented large-scale comparison across different audio modalities. The results from this benchmarking reveal significant deficiencies in several key areas:
- Acoustic Grounding: Many models struggle to accurately link their outputs to the acoustic features present in the input audio.
- Temporal Reasoning: The ability to understand and process audio over time is a significant challenge for current LALMs.
- Music Attribute Understanding: There are notable shortcomings in models’ capacities to accurately interpret and describe attributes of music.
Conclusion
The development of HalluAudio marks a significant milestone in the evaluation of hallucination detection within audio-language models. By providing a comprehensive and large-scale benchmark, it paves the way for further advancements in the reliability and robustness of LALMs. As the field continues to evolve, addressing the identified deficiencies will be crucial in enhancing the performance and applicability of these models in real-world scenarios.
For those interested in further details, the full paper can be accessed on arXiv under the identifier arXiv:2604.19300v1.