{
  "@context": [
    "https://schema.org",
    {
      "rai": "http://mlcommons.org/croissant/RAI/",
      "prov": "http://www.w3.org/ns/prov#",
      "sc": "https://schema.org/"
    }
  ],
  "@type": "Dataset",
  "name": "blab_long_audio",
  "description": "\n\t\n\t\t\n\t\tBLAB: Brutally Long Audio Bench\n\t\n\n\n\t\n\t\t\n\t\tDataset Summary\n\t\n\nBrutally Long Audio Bench (BLAB) is 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 Youtube audio clips, each paired with human-annotated, text-based natural language questions and answers. Our audio data were collected from permissively… See the full description on the dataset page: https://huggingface.co/datasets/oreva/blab_long_audio.",
  "alternateName": [
    "oreva/blab_long_audio",
    "BLAB (Brutally Long Audio Bench)"
  ],
  "creator": {
    "@type": "Person",
    "name": "Orevaoghene Ahia",
    "url": "https://huggingface.co/oreva"
  },
  "keywords": [
    "audio-text-to-text",
    "original",
    "cc-by-4.0",
    "1K<n<10K",
    "Audio",
    "arxiv:2505.03054",
    "🇺🇸 Region: US",
    "speech",
    "audio",
    "speech-llm",
    "audio-lm",
    "long-audio",
    "spoken-language-understanding"
  ],
  "license": "https://choosealicense.com/licenses/cc-by-4.0/",
  "url": "https://huggingface.co/datasets/oreva/blab_long_audio",
  "rai:dataLimitations": "BLAB is currently English-only, sourced from English-language YouTube content, and results may not generalize to multilingual or non-English audio understanding. Audio is sourced exclusively from YouTube, which may not represent all real-world audio domains, content skews toward public-facing media such as podcasts, debates, and interviews, and may underrepresent private or domain-specific audio such as clinical recordings. Accent and gender are not controlled or balanced across samples and may reflect biases present in publicly available YouTube content, though we made efforts to handpick audio files with diverse speaker profiles.\n\nBLAB is an evaluation benchmark only . it is discouraged to use it for model training. BLAB should not be used to evaluate general multimodal audio understanding beyond the specific tasks and domains it covers.",
  "rai:dataBiases": "Audio content is sourced exclusively from YouTube, which overrepresents publicly available, permissively licensed media content such as debate podcasts, comedy shows, interviews, and panel discussions. This may not reflect the full diversity of real-world audio contexts such as private conversations, clinical settings, or spontaneous speech.\nSpeaker demographics including age, accent, dialect, and gender are not systematically controlled or balanced. While YouTube hosts content from speakers across many countries and linguistic backgrounds, the English-language filter applied during data collection means speakers are predominantly proficient English speakers, which may underrepresent non-native English speakers, speakers with speech impairments, elderly speakers, and regional dialects.\nGround truth for word localization tasks is derived from automated forced alignment tools (WhisperX) with human verification, which may introduce minor systematic errors for overlapping speech, accented speech, or low-quality recordings\nModels may perform differently on BLAB than on real-world audio that differs in speaker demographics, recording conditions, or content domain, and BLAB performance should not be taken as a direct predictor of performance in deployment settings that differ substantially from the benchmark's distribution.",
  "rai:personalSensitiveInformation": "BLAB contains audio from publicly available, Creative Commons-licensed YouTube content. Personal attributes that may be incidentally present include gender and age (audible from speaker voices), geography and culture (inferable from accents and content topics), and political or religious beliefs (present in debate and interview content). No health or medical data, socio-economic status, or experience and seniority information is deliberately collected or annotated. All data was sourced in accordance with YouTube's terms of service and applicable privacy guidelines.",
  "rai:dataUseCases": "BLAB is designed to measure long-form audio reasoning capabilities in multimodal language models specifically, the ability to perform temporal localization, speaker counting, emotion interpretation, and duration estimation over audio recordings spanning 15 minutes to 2 hours. It is intended to represent the gap between short-form audio evaluation and the reasoning demands of real-world long-form audio.",
  "rai:dataSocialImpact": "BLAB is intended to advance the development of audio AI systems with positive applications in accessibility, healthcare, legal, and media domains. However, long-form audio understanding capabilities particularly speaker tracking and emotion recognition carry risks of misuse for surveillance or non-consensual monitoring. All audio is sourced from Creative Commons-licensed YouTube content, and the benchmark is released for research and evaluation purposes only. We explicitly discourage use of BLAB for training surveillance systems or making consequential decisions about individuals, and encourage users to complement benchmark evaluation with fairness audits on diverse speaker populations.",
  "rai:hasSyntheticData": false,
  "prov:wasDerivedFrom": [
    {
      "@id": "youtube.com",
      "prov:label": "YouTube"
    }
  ],
  "prov:wasGeneratedBy": [
    {
      "@type": "prov:Activity",
      "prov:type": {
        "@id": "https://www.wikidata.org/wiki/Q4929239"
      },
      "prov:label": "Data collection",
      "sc:description": "Audio files were handpicked from YouTube, filtering for Creative Commons-licensed content across diverse genres including interviews, podcasts, debates, and media content, prioritizing complexity, speaker diversity, and task relevance. Preprocessing involved for localization tasks and tool use analysis was done using WhisperX and  Pyannote. Human annotation was conducted by the paper authors across all tasks — annotators verified extracted spans, entity types, and timestamp alignments, with disagreements resolved by a third annotator where applicable. All question-answer pairs were either handcrafted and verified."
    }
  ]
}