Abstract
Large Audio Language Models (LALMs) are transforming AI by processing and generating human language directly from audio. As these models proliferate in real-world applications, it becomes critical to evaluate their performance to ensure equitable and safe use across diverse linguistic and cultural contexts. We present the first comprehensive study of cultural bias in LALMs, extending text-based harm frameworks to the audio modality to analyze how linguistic diversity influences model behavior and uncover challenges in interpreting audio nuances. To address this, we introduce the Audio Cultural Intelligence Dataset (ACID), a multilingual audio–text benchmark spanning 1,315 hours across diverse languages and cultural contexts, and we conduct a systematic evaluation of 10 open-source and two closed-source models. Our results reveal substantial performance disparities across languages and cultural settings and show that biases manifest distinctly when models process audio inputs. These findings highlight the need to evaluate LALMs not only for technical accuracy but also for fair and culturally sensitive behavior, motivating the development of inclusive datasets and culturally aware training practices for safer and more equitable audio language models.
Introduction
Large Audio Language Models (LALMs) such as SALMONN and Qwen-Audio represent the next frontier in human-AI interaction. Unlike traditional pipelines that transcribe speech to text before processing, LALMs "hear" audio directly. This capability allows them to potentially capture emotion, tone, and prosody—elements often lost in text. However, as these models are deployed globally, a critical question arises: Do they understand the cultural nuances of the voices they hear, or are they projecting Western-centric biases onto a multilingual world?
To address this, researchers from IIT Jodhpur and FLAME University introduced the Audio Cultural Intelligence Dataset (ACID), the first comprehensive benchmark designed to measure cultural safety and alignment in LALMs.
The ACID Benchmark
The ACID benchmark addresses the "cultural misalignment" inherent in models trained predominantly on high-resource languages and Western datasets. Speech is a rich medium where meaning can shift based on accent, dialect, or cultural context. Without culturally aware evaluation, models may provide answers that are technically "safe" by Western standards but offensive or inaccurate within local frameworks.
ACID consists of 1,315 hours of audio divided into three specialized subsets:
- Set-A (Cultural Harm): 11,620 samples for evaluating single-turn interactions.
- Set-B (Contextual Sensitivity): 77,860 samples focused on multi-turn conversational sequences to observe how biases evolve during dialogue.
- Set-C (Alignment & Preference): 300,000 samples designed for fine-tuning models using preference-based learning (e.g., DPO).
The dataset covers 10 languages (Arabic, Bengali, English, French, Gujarati, Hindi, Russian, Telugu, Turkish, and Vietnamese) across 12 societal dimensions, including Religion, Ethics, Security, and Migration.
Methodology
The researchers developed a rigorous pipeline to create ACID, adapting the CulturalKaleidoscope framework to identify 1,162 culturally sensitive prompts. These were translated using state-of-the-art models (NLLB and Opus-MT) and synthesized into high-quality audio using MMS-TTS. The evaluation suite included 10 open-source models (such as MERaLiON and LTU-AS) and 2 closed-source models (Gemini-1.5 Flash and GPT-4o Mini), utilizing metrics for relevance, sentiment, and safety.
Qualitative Analysis of Model Failures
The evaluation revealed several critical failure modes that highlight the gap between technical accuracy and cultural intelligence:
- Transcription as a Default: When faced with complex or sensitive questions, many models "gave up" on reasoning, simply transcribing the audio input rather than answering the prompt.
- Safety vs. Irrelevance: Some models achieved high safety scores by outputting generic, irrelevant phrases like "person speaking" or "door." While technically avoiding harm, these responses render the model functionally useless.
- Dangerous Compliance: In some instances, models exhibited a lack of safety alignment by providing definitions for harmful premises (e.g., "degenerate neighbors") rather than refusing the biased prompt.
Quantitative Results and Disparities
- Performance Leaders: MERaLiON emerged as the strongest open-source performer in relevance (0.35–0.46), though no model exceeded a score of 0.50, indicating a significant gap in true conversational comprehension.
- Closed-Source Superiority: GPT-4o Mini and Gemini-1.5 Flash outperformed open-source baselines in relevance and sentiment. However, GPT-4o Mini’s high sentiment scores often reflected "context-appropriate politeness" rather than deep cultural understanding.
- The Digital Divide: There were substantial performance disparities across languages. Models generally performed better in English and French than in lower-resource languages like Telugu or Gujarati, underscoring the need for more inclusive training data.
Conclusion
While the ACID benchmark is a major step forward, the low absolute relevance scores across all tested models suggest that current LALM architectures struggle to map audio signals to complex cultural reasoning. Future work must focus on culturally aware training, audio-specific alignment guardrails that understand prosody, and expanding language coverage to include more diverse linguistic families and dialects. By releasing ACID, the researchers aim to catalyze the development of AI that truly understands the cultural context of every voice.
Summary updated