vocal-burst-classification-v2
收藏资源简介:
Vocal Burst Classification V2(简称V2)是由LAION发布的数据集,用于训练Vocal Burst Classifier V2模型。该数据集是一个单标签分类数据集,包含83个类别(82种非语音人声爆发类别,加上一个“no_burst”类别,索引为82)。数据来源包括两个部分:一是通过合成文本到音频模型(如DramaBox)生成的音频片段,二是来自先前版本的vocal-bursts-clean音频。标签由两个标注器产生:Gemini Flash 2.5对improved_synthetic_vocal_burts片段进行验证,Gemini 3.6 Flash对vocal-bursts-clean进行盲重标注。数据集包含多个子集:train_existing.npz(70176个样本,768维VoiceCLAP嵌入)、train_gemini.npz(25245个样本,768维嵌入,包含src字段标识来源)、val.npz(验证集嵌入,包含pure和composite两种类型),以及clean_audio配置(18329个原始MP3音频,字段包括id、audio、gemini_36_label、original_noisy_label和source)。此外还提供了类列表(classes.json)、类描述(class_descriptions.json)、源计数(source_counts.json)以及训练、推理和特征提取脚本。注意,手/身体冲击类(如Finger Snaps、Hand Scratching Head等)虽保留在类列表中用于向后兼容,但已被折叠到no_burst类别,训练样本中计数为0。该数据集适用于音频分类任务,特别是人声爆发检测与分类。
Vocal Burst Classification V2 (V2) is a dataset released by LAION for training the Vocal Burst Classifier V2 model. It is a single-label classification dataset with 83 categories (82 non-speech vocal burst categories plus a no_burst category with index 82). The data sources include two parts: audio clips generated by synthetic text-to-audio models (such as DramaBox) and audio from the previous version of vocal-bursts-clean. Labels are produced by two annotators: Gemini Flash 2.5 validates the improved_synthetic_vocal_bursts clips, and Gemini 3.6 Flash performs blind re-labeling on vocal-bursts-clean. The dataset contains multiple subsets: train_existing.npz (70,176 samples with 768-dimensional VoiceCLAP embeddings), train_gemini.npz (25,245 samples with 768-dimensional embeddings and a src field indicating source), val.npz (validation set embeddings with pure and composite types), and a clean_audio configuration (18,329 original MP3 audios with fields including id, audio, gemini_36_label, original_noisy_label, and source). Additionally, class list (classes.json), class descriptions (class_descriptions.json), source counts (source_counts.json), and training, inference, and feature extraction scripts are provided. Note that hand/body impact classes (e.g., Finger Snaps, Hand Scratching Head) are retained in the class list for backward compatibility but have been folded into the no_burst category, with zero counts in training samples. This dataset is suitable for audio classification tasks, especially vocal burst detection and classification.
Vocal Burst Classification V2 数据集概述
基本信息
- 数据集名称:
laion/vocal-burst-classification-v2 - 许可证:Apache-2.0
- 任务类型:音频分类(audio-classification)
- 配置:仅包含
clean_audio配置,分为训练集(18,329 个样本) - 数据集大小:约 178.9 MB
数据集内容
该数据集是 Vocal Burst Classifier V2 的训练语料库,包含:
- 83 类发声爆发分类(82 类非语音人类发声 +
no_burst类别) - 预计算的 VoiceCLAP-commercial 嵌入特征(768 维)
- 原始音频文件(
vocal-bursts-clean,MP3 格式)
数据集结构
主要包含以下文件:
embeddings/train_existing.npz— 先前发布的训练嵌入特征embeddings/train_gemini.npz— 新的 V2 Gemini 训练集(25,245 个样本)embeddings/val.npz— 官方验证集嵌入特征clean_audio配置 — 18,329 个原始 MP3 音频文件classes.json、class_descriptions.json、source_counts.jsoninference.py、train.py、extract_features.py
数据字段
每个样本包含:
id:字符串标识符audio:音频数据gemini_36_label:Gemini 3.6 标注的标签original_noisy_label:原始标注(含噪声)source:数据来源
标注来源
- Gemini Flash 2.5 — 验证改进的合成声音片段(
src == "improved") - Gemini 3.6 Flash — 对
vocal-bursts-clean进行盲重新标注(src == "clean"),作为训练目标
关键特点
- 原始标注噪声较大,Gemini 3.6 与原始标签一致率仅为 5.8%
- 手部/身体冲击类声音(如手指弹响、拍手等)已合并到
no_burst类别 - 训练集中各类别样本数从 0 到 1,608 不等,分布不均衡
模型性能
| 模型 | 验证集准确率 | 纯集(1170) | 混合集(770) |
|---|---|---|---|
| V2 混合 emb-MLP(发布版本) | 58.1% | 70.4% | 39.4% |
| 先前研究:混合 emb-MLP | 58.61% | 72.05% | 38.18% |
类别分布
类别涵盖笑声、哭泣、呼吸、叹气、喘息、呻吟、咕哝、喉咙声、哼唱、咳嗽、口哨、口唇声、舌头点击、饮食声、尖叫等 16 大类,各类别样本数量从 0 到 1,608 不等。




