BLAB: Brutally Long Audio Bench

TL;DR

BLAB tackles the gap in long-form audio understanding by introducing a large-scale, eight-task benchmark (localization, counting, emotion, and duration) with samples from $15$ minutes to $2$ hours, totaling $833+$ hours. It systematically curates data from Creative Commons YouTube content and evaluates state-of-the-art and open-weight audio LMs, revealing substantial difficulties in long-form reasoning, precise localization, and counting as duration grows. The study provides a detailed methodology, prompts, and metrics, and demonstrates that even powerful models like Gemini 2.0 Pro and GPT-4o struggle significantly on these tasks, highlighting the need for robust long-context, open-source multimodal models. The work sets a foundation for future research in long-form audio processing, emphasizing reproducibility and transparency through public data, code, and model generations.

Abstract

Developing large audio language models (LMs) capable of understanding diverse spoken interactions is essential for accommodating the multimodal nature of human communication and can increase the accessibility of language technologies across different user populations. Recent work on audio LMs has primarily evaluated their performance on short audio segments, typically under 30 seconds, with limited exploration of long-form conversational speech segments that more closely reflect natural user interactions with these models. We introduce Brutally Long Audio Bench (BLAB), a challenging long-form audio benchmark that evaluates audio LMs on localization, duration estimation, emotion, and counting tasks using audio segments averaging 51 minutes in length. BLAB consists of 833+ hours of diverse, full-length audio clips, each paired with human-annotated, text-based natural language questions and answers. Our audio data were collected from permissively licensed sources and underwent a human-assisted filtering process to ensure task compliance. We evaluate six open-source and proprietary audio LMs on BLAB and find that all of them, including advanced models such as Gemini 2.0 Pro and GPT-4o, struggle with the tasks in BLAB. Our comprehensive analysis reveals key insights into the trade-offs between task difficulty and audio duration. In general, we find that audio LMs struggle with long-form speech, with performance declining as duration increases. They perform poorly on localization, temporal reasoning, counting, and struggle to understand non-phonemic information, relying more on prompts than audio content. BLAB serves as a challenging evaluation framework to develop audio LMs with robust long-form audio understanding capabilities.

Figures (5)

• Figure 1: Overview of BLAB, designed to test true multimodal understanding abilities of audio LMs. It contains eight distinct audio tasks across four categories, namely localization, counting, emotion, and duration estimation. $^\dag$ All images are designed by https://www.freepik.com/Freepik .
• Figure 2: Comparison of predicted and ground truth values for speaker number estimation and entire duration tasks
• Figure 3: Comparison of long audio and short audio results across Gemini Models
• Figure 4: Performance comparison when the original audio input is replaced with silence or Gaussian noise. Since the entire input is noisy, the ground truth label is 0
• Figure 5: Placing a 30-second clean audio clip at different points within a long, noisy audio input impacts speaker number estimation performance. Proprietary models like Gemini perform better when the clean clip is positioned at the beginning or end of the noisy audio.