International Multi-Speaker AUDIO (+Video) Dataset with Stems, Diarized Transcripts, Scene ...

The ACNetwork Multilingual Conversational Media Dataset delivers 280K hours of UCG audio and video data across more than 20 languages: including English (70,665 hrs), Portuguese (38,282 hrs), Spanish (26,928 hrs), Russian (10,182 hrs), French (8,127 hrs), German (5,057 hrs), Italian (3,202 hrs), Japanese (2,228 hrs), and Korean (2,147 hrs)

Each file features high-fidelity audio with isolated stems, human-verified transcripts at word and utterance level, speaker diarization, emotion and tone labels, and scene metadata for video segments. All content is fully rights-cleared and indemnified for AI/LLM training and commercial reuse, making it ideal for speech-to-text, multimodal retrieval, cross-lingual fine-tuning, and emotion classification.

The dataset spans sports, news, entertainment, gaming, true crime, and more, with new content added monthly to ensure domain diversity and freshness. Delivery is via Google Drive or AWS S3, and data is provided in multiple formats (JSON, SRT, VTT, TXT).

This is the definitive global multilingual sports and talk dataset for enterprise-scale AI and LLM training - combining audio and video sources at unmatched depth and breadth. Enterprise licensing only - contact ACNetwork for commercial terms or expanded sample access.

搜集汇总

背景与挑战

背景概述

该数据集提供超过28万小时的多语言音频和视频内容，涵盖20多种语言，并包含高保真音频、分离音轨、人工转录、说话人日志、情感标签及场景元数据等丰富标注。数据覆盖体育、新闻、娱乐等多个领域，每月持续更新，适用于语音识别、多模态检索等AI训练场景，仅限企业级许可使用。