vocalburst-locator-synth-data
收藏资源简介:
Vocal Burst Locator - Synthetic Training Data是一个用于训练音频事件检测模型的合成数据集,专门针对人声爆发(如笑声、咳嗽、喷嚏、叹息、喘息、哭泣、尖叫等)的定位任务。该数据集由LAION组织创建,用于训练基于Whisper的vocalburst-locator模型,以检测和定位音频中的人声爆发事件。数据集包含33,012个音频样本,总计约8.5GB,分为训练集(32,712个样本,~8.4GB)和验证集(300个样本,~79MB)。每个样本包括单声道MP3音频文件(44.1kHz采样率,96kbps比特率,持续时间3-30秒)和对应的JSON标签文件,标签文件包含人声爆发事件的时间戳、音频总时长、背景类型(音乐、音效、音乐+音效或静音)以及人声爆发数量。数据集设计平衡:50%的样本为负样本(无人声爆发),50%为正样本(包含至少一个人声爆发),人声爆发数量分布以1-2个爆发为主,整个数据集共包含31,360个人声爆发事件。背景音频类型分布均匀:音乐、音效和音乐+音效各占约31.7%,静音背景占5.0%。音频持续时间遵循贝塔分布,偏向较长的片段(20-30秒)。数据集通过合成方法生成,生成流程包括从源数据集获取素材、随机选择背景类型和持续时间、混合背景音频并放置人声爆发片段,以及应用数据增强(如低通滤波和加性噪声)。所有人声爆发源片段在数据集中恰好出现两次。数据集采用Apache 2.0许可证,适用于音频事件检测、声音场景分析和机器学习模型训练等任务。
Vocal Burst Locator - Synthetic Training Data is a synthetic dataset designed for training audio event detection models, specifically focusing on the localization task of vocal bursts such as laughter, coughing, sneezing, sighing, gasping, crying, and screaming. Created by the LAION organization, this dataset is intended for training the Whisper-based vocalburst-locator model to detect and localize vocal burst events in audio. The dataset consists of 33,012 audio samples totaling approximately 8.5GB, split into a training set (32,712 samples, ~8.4GB) and a validation set (300 samples, ~79MB). Each sample includes a monaural MP3 audio file with a 44.1kHz sampling rate, 96kbps bitrate, and duration ranging from 3 to 30 seconds, alongside a corresponding JSON label file. The label file contains the timestamps of vocal burst events, total audio duration, background type (music, sound effects, music + sound effects, or silence), and the number of vocal bursts. The dataset is balanced: 50% of the samples are negative samples (containing no vocal bursts) and 50% are positive samples (containing at least one vocal burst). The distribution of vocal burst counts is dominated by 1 or 2 bursts per sample, and the entire dataset contains a total of 31,360 vocal burst events. The background audio types are evenly distributed: music, sound effects, and music + sound effects each account for approximately 31.7% of the samples, while silent backgrounds make up 5.0%. The audio duration follows a beta distribution, with a bias towards longer segments (20–30 seconds). The dataset is generated through a synthetic pipeline that includes acquiring source materials from existing datasets, randomly selecting background types and durations, mixing background audio and inserting vocal burst segments, and applying data augmentation techniques such as low-pass filtering and additive noise. Notably, every vocal burst source segment appears exactly twice within the dataset. The dataset is licensed under Apache 2.0, and is suitable for applications including audio event detection, sound scene analysis, and machine learning model training.
数据集概述
Vocal Burst Locator - Synthetic Training Data 是一个合成的音频数据集,专用于训练语音爆发事件检测与定位模型。数据集包含 33,012 个音频-标签对,总大小约 8.5 GB。
数据集规模与划分
| 划分 | 样本数 | 大小 |
|---|---|---|
| 训练集 | 32,712 | ~8.4 GB |
| 验证集 | 300 | ~79 MB |
| 总计 | 33,012 | ~8.5 GB |
数据格式
每个样本包含两个文件:
sample_XXXXX.mp3— 单声道音频音景(3–30 秒,44.1 kHz,96 kbps MP3)sample_XXXXX.json— 包含语音爆发时间戳和元数据的标签文件
标签文件 JSON 结构:
| 字段 | 类型 | 描述 |
|---|---|---|
events |
list | 语音爆发片段,包含 start_time 和 end_time(秒) |
duration_sec |
float | 音频总时长(秒) |
bg_type |
string | 背景类型:"music"、"sfx"、"music+sfx" 或 "silence" |
n_vocal_bursts |
int | 该片段中的语音爆发数量 |
当 n_vocal_bursts: 0 时,events 列表为空(负样本)。
数据集构成
类别平衡
- 50% 负样本(无语音爆发):16,506 个样本
- 50% 正样本(至少 1 个语音爆发):16,506 个样本
每个片段中语音爆发数量分布:
| 语音爆发数 | 样本数 |
|---|---|
| 0 | 16,506 |
| 1 | 6,690 |
| 2 | 6,063 |
| 3 | 2,747 |
| 4 | 771 |
| 5 | 201 |
| 6+ | 34 |
整个数据集共有 31,360 个语音爆发事件。
背景类型分布
| 背景类型 | 样本数 | 占比 |
|---|---|---|
| 音乐 | 10,464 | 31.7% |
| 音效 | 10,465 | 31.7% |
| 音乐 + 音效 | 10,433 | 31.6% |
| 静音 | 1,650 | 5.0% |
时长分布
片段时长由 Beta 分布生成:Beta(3.0, 1.2) * 27 + 3,范围为 3–30 秒,大部分片段集中在 20–30 秒。
数据生成方法
数据集通过将语音爆发音频片段混合到背景音频上合成生成。
1. 来源音频
| 来源 | 数据集 | 数量 | 描述 |
|---|---|---|---|
| 语音爆发 | laion/improved_synthetic_vocal_burts | 15,680 | 合成的语音爆发片段 |
| 音乐 | laion/captioned-ai-music-snippets | 5,000 | AI 生成的音乐片段 |
| 音效 (AudioSet) | mitermix/audioset-with-grounded-captions | 5,000 | AudioSet 音效 |
| 音效 (AudioSnippets) | mitermix/audiosnippets_small_with_detailed_annotation | 3,000 | 多样化的音频片段 |
2. 音景生成流程
- 从 Beta 分布中选取时长(3–30 秒)
- 根据目标分布选择背景类型
- 随机选择和混合来源音频创建背景
- 对于正样本(50%),在随机位置放置 1–5+ 个语音爆发片段,音量为 10–100%(25% 概率重叠)
- 可选增强(15% 样本):低通滤波和/或加性噪声
- 导出为单声道 MP3 和 JSON 标签文件
每个语音爆发源片段在整个数据集中恰好出现 两次。
3. 训练/验证划分
- 300 个样本用于验证
- 剩余 32,712 个用于训练
- 划分是确定性的(种子 = 2024)
文件结构
vocalburst-locator-synth-data/ train/ shard_00/ # 样本 00000–03999(4,000 个样本) shard_01/ # 样本 04000–07999 ... shard_08/ # 样本 32000–32711(712 个样本) val/ # 300 个样本(扁平目录) manifest.json # 数据集生成元数据
生成参数
| 参数 | 值 |
|---|---|
| 目标采样率 | 44,100 Hz |
| MP3 比特率 | 96 kbps |
| 时长分布 | Beta(3.0, 1.2) * 27 + 3 |
| 负样本比例 | 50% |
| VB 音量范围 | 10–100% |
| VB 重叠概率 | 25% |
| 质量增强概率 | 15% |
| 随机种子 | 2024 |
许可证
Apache 2.0




