StreamAudio-2M
收藏资源简介:
StreamAudio-2M 是一个专为音频大语言模型和音频智能体训练设计的大规模流式音频数据集。其核心特点在于以“流”为单位组织数据,每个流是一个共享统一模式的音频回合序列。数据集总计包含约 228 万个独特的音频片段,这些片段被系统地组织成六个任务特定的子集,每个子集针对不同的音频处理场景。这些子集包括:1. 流式音频理解(90,738 行),包含来自 AudioSet 和 FMA 的音频蒙太奇,用于字幕生成、选择和开放式问答任务;2. 实时自动语音识别(28,109 行),包含来自 CommonVoice、GigaSpeech、LibriSpeech 和 VoxPopuli 的 ASR 音频流,其中助手回复为文本转录;3. 流式翻译(11,712 行),包含来自 CoVoST2 的英语到中文语音翻译音频流,助手回复为中文翻译;4. 语音聊天(196,064 行),包含多轮口语对话,每回合都有语音形式的用户和助手交互;5. 主动响应(4,519 行),包含室内/室外场景下的主动响应音频流,来源包括 ElevenLabs 和 AudioX;6. 环境感知音频智能体(50,035 行),包含相同场景下的音频蒙太奇,时长在 90 到 180 秒之间。数据集采用统一的结构化模式:顶层包含流 ID、场景类型、音频类别、回合数、总时长和回合列表等字段。每个回合则包含用户音频、助手音频、情感、关键词、场景类型、音频类别、话题类型、音频文件路径、数据来源以及详细的音频统计信息(如时长、采样率、响度、动态范围等)。音频文件以未压缩的 tar 分片形式提供,解压后可按指定相对路径访问。该数据集适用于多种音频相关的任务,包括音频分类、自动语音识别、语音翻译和音频到音频的生成等,旨在支持流式、多轮交互的音频理解和生成模型的开发。
StreamAudio-2M is a large-scale streaming audio dataset designed for training audio large language models and audio agents. Its core feature is organizing data in streams, where each stream is a sequence of audio turns sharing a unified pattern. The dataset contains approximately 2.28 million unique audio clips, systematically organized into six task-specific subsets, each targeting different audio processing scenarios. These subsets include: 1. Streaming Audio Understanding (90,738 rows), containing audio montages from AudioSet and FMA for caption generation, selection, and open-ended QA tasks; 2. Real-time Automatic Speech Recognition (28,109 rows), containing ASR audio streams from CommonVoice, GigaSpeech, LibriSpeech, and VoxPopuli, with assistant responses as text transcriptions; 3. Streaming Translation (11,712 rows), containing English-to-Chinese speech translation audio streams from CoVoST2, with assistant responses as Chinese translations; 4. Voice Chat (196,064 rows), containing multi-turn spoken dialogues with user and assistant interactions in audio form per turn; 5. Proactive Response (4,519 rows), containing proactive response audio streams in indoor/outdoor scenarios, sourced from ElevenLabs and AudioX; 6. Environment-aware Audio Agent (50,035 rows), containing audio montages in the same scenarios, with durations between 90 and 180 seconds. The dataset adopts a unified structured schema: the top level includes fields such as stream ID, scene type, audio category, number of turns, total duration, and turn list. Each turn contains user audio, assistant audio, emotion, keywords, scene type, audio category, topic type, audio file path, data source, and detailed audio statistics (e.g., duration, sample rate, loudness, dynamic range). Audio files are provided in uncompressed tar shards, which can be accessed via specified relative paths after extraction. This dataset is suitable for various audio-related tasks, including audio classification, automatic speech recognition, speech translation, and audio-to-audio generation, aiming to support the development of streaming, multi-turn interactive audio understanding and generation models.
数据集概览
数据集名称:StreamAudio-2M
许可证:CC-BY-4.0
语言:英语、中文
任务类型:音频分类、自动语音识别、翻译、音频到音频
标签:音频、流式、音频理解、ASR、语音翻译、语音聊天
StreamAudio-2M 是一个大规模流式音频数据集,专为音频大语言模型和音频智能体训练设计。每个数据项是一个流(stream),即一系列音频片段按统一模式组织。数据集包含约 228 万条独特的音频片段,划分为 6 个任务子集。
子集说明
| 子集名称 | 行数 | 描述 |
|---|---|---|
| Stream_Audio_Understanding | 90,738 | 音频理解片段合集(来源于 AudioSet / FMA),包含描述、选择题和开放性问答 |
| Real_time_ASR | 28,109 | 自动语音识别片段流(来源于 CommonVoice / GigaSpeech / LibriSpeech / VoxPopuli);assistant 字段为文本转写 |
| Stream_translation | 11,712 | 英文到中文的语音翻译片段流(来源于 CoVoST2);assistant 字段为中文翻译 |
| Voice_chatting | 196,064 | 多轮口语对话,每轮包含语音形式的 user 和 assistant |
| Proactive_respond | 4,519 | 室内/室外主动响应流(来源于 ElevenLabs / AudioX) |
| Environment_awared_audio_agent | 50,035 | 同一场景的音频合集(时长 90–180 秒) |
数据模式
顶层字段:
id:唯一标识符stream_scene_type:所有轮次一致的场景类型,否则为mixedstream_audio_category:所有轮次一致的音频类别,否则为mixednum_turns:流中的轮次数total_duration_s:总时长(秒)turns:轮次列表,每个轮次包含以下字段:user:用户输入assistant:助手输出emotion:情感keywords:关键词scene_type:场景类型,取值范围包括Driving、Traffic、Home Smart、Cooking、Fitness、Travel、Office、othersaudio_category:音频类别topic_type:话题类型audio_path:音频文件路径source:数据来源audio_stats:音频统计信息,包含duration_s、sample_rate、channels、mean、std、abs_mean、rms_db、peak_db、crest_factor、dynamic_range_db、zero_crossing_rate
音频文件
音频以未压缩的 tar 分片形式存储在 audio_tars/ 目录下。下载所有分片并解压到数据集根目录后,可恢复 audio/ 目录树。数据集中每个轮次的 audio_path 字段为相对于 audio/ 的路径,例如 audio/asr/0012/asr_0012345.wav。
解压命令示例: bash cat audio_tars/*.tar | tar -xf -
或单独解压每个分片。




