PC-Mix
收藏资源简介:
PC-Mix(部分伪造数据集,带语音-环境声音受控混合)是一个用于音频反欺骗研究的开源数据集,专注于部分伪造音频的检测。该数据集通过将来自PartialSpoof v1.2的语音样本与一个自建的部分伪造环境声音池进行受控混合,创建了具有不同语音和环境声音真实性组合的音频样本。原始的真实录音来源于VGGSound和SBCSAE数据集。数据集提供话语级和帧级双重标注:话语级包含5个类别,包括原始真实录音和四种构造混合类型(真实+真实、伪造+真实、真实+伪造、伪造+伪造);帧级标注针对构造混合样本,将每一帧划分为4类(两者均真实、仅语音伪造、仅环境声音伪造、两者均伪造)。数据规模方面,训练集包含25,380个混合样本和12,155个原始样本;评估集包含5个子集(E0-E4),总计约71,000个混合样本和17,809个原始样本,用于测试模型在不同分布外条件下的泛化能力。数据集文件组织为两个主要目录:Mixed_soundsV1.2包含混合音频文件及其标签,按训练/评估分割存储;Partialspoof_background包含源环境声音数据。每个分割都包括协议文件、WAV音频文件和分层标签。由于许可证限制,部分原始音频片段未直接分发,但提供了完整的重建脚本和协议。PC-Mix适用于音频分类、反欺骗、部分伪造检测、音频取证等任务,特别适合研究复杂声学场景中多成分真实性的细粒度分析。
PC-Mix (Partial Spoof Dataset with Controlled Speech–Environmental Sounds Mixing) is an open-source dataset for audio anti-spoofing research, focusing on the detection of partially spoofed audio. It is constructed by controlled mixing of speech samples from PartialSpoof v1.2 with a self-built partially spoofed environmental sound pool, creating audio samples with varying combinations of authenticity in speech and environmental sounds. The original genuine recordings are sourced from the VGGSound and SBCSAE datasets. The dataset provides dual-level annotations: utterance-level includes 5 categories, covering original genuine recordings and four constructed mixture types (genuine+genuine, spoofed+genuine, genuine+spoofed, spoofed+spoofed); frame-level annotations target constructed mixture samples, classifying each frame into 4 categories (both genuine, only speech spoofed, only environmental sound spoofed, both spoofed). In terms of data scale, the training set contains 25,380 mixture samples and 12,155 original samples; the evaluation set comprises 5 subsets (E0-E4), totaling approximately 71,000 mixture samples and 17,809 original samples, designed to test model generalization under various out-of-distribution conditions. The dataset files are organized into two main directories: Mixed_soundsV1.2 contains mixed audio files and their labels, stored by training/evaluation splits; Partialspoof_background contains source environmental sound data. Each split includes protocol files, WAV audio files, and hierarchical labels. Due to licensing restrictions, some original audio segments are not directly distributed, but complete reconstruction scripts and protocols are provided. PC-Mix is suitable for tasks such as audio classification, anti-spoofing, partial spoof detection, and audio forensics, particularly for fine-grained analysis of multi-component authenticity in complex acoustic scenarios.
PC-Mix 数据集概述
基本信息
- 名称: PC-Mix (Partial Spoof Dataset with Controlled Speech–Environmental Sounds Mixing)
- 语言: 英语
- 任务: 音频分类
- 标签: 音频、反欺骗、部分欺骗音频、环境声音
- 许可: 非商用研究目的
数据来源
| 组件 | 来源 | 描述 |
|---|---|---|
| 语音 | PartialSpoof v1.2 | 带有部分欺骗标注的语音样本 |
| 环境音 | 自策部分欺骗环境音池 | 带有受控真实/欺骗组件标注的环境音 |
| 原始录音 | VGGSound 和 SBCSAE | 无人工语音-环境混合的自然录音 |
数据集划分
训练集
| 划分 | 混合样本数 | 原始样本数 | 环境音来源 | 事件来源 | 融合方式 |
|---|---|---|---|---|---|
| Train | 25,380 | 12,155 | SONYC | 60% AudioLDM2;40% UrbanSound8K (0-1秒) | Ducking Overlay |
评估集
| 子集 | 混合样本 | 原始样本 | 环境音 | 事件 | 融合方式 |
|---|---|---|---|---|---|
| E0 Baseline | 17,812 | 17,809 | SONYC | 60% AudioLDM2;40% FSD50K (0-1秒) | Ducking Overlay |
| E1 Generator OOD | 14,248 | — | SONYC | AudioGen | Ducking Overlay |
| E2 Fusion OOD | 14,248 | — | SONYC | AudioLDM2 | Energy Matching + Crossfade (20-80毫秒) |
| E3 Environmental Sounds OOD | 17,809 | — | DEMAND | AudioLDM2 | Ducking Overlay |
| E4 Noise OOD | 7,125 | — | WHAM Noise | AudioLDM2 | Ducking Overlay |
标注体系
话语级类别(5类)
| 类别 | 名称 | 语音 | 环境音 | 描述 |
|---|---|---|---|---|
| 0 | Original | — | — | 自然录音,无人工混合 |
| 1 | Bona fide + Bona fide | 真实 | 真实 | 语音和环境音均为真实 |
| 2 | Spoofed + Bona fide | 部分欺骗 | 真实 | 语音部分欺骗,环境音真实 |
| 3 | Bona fide + Spoofed | 真实 | 部分欺骗 | 语音真实,环境音部分欺骗 |
| 4 | Spoofed + Spoofed | 部分欺骗 | 部分欺骗 | 语音和环境音均含部分欺骗区域 |
帧级类别(4类)
| 类别 | 名称 | 语音 | 环境音 | 描述 |
|---|---|---|---|---|
| 0 | Bona fide frame | 真实 | 真实 | 两组件均未被欺骗 |
| 1 | Speech-spoofed frame | 欺骗 | 真实 | 仅语音组件被局部欺骗 |
| 2 | Environmental-sound-spoofed frame | 真实 | 欺骗 | 仅环境音组件被局部欺骗 |
| 3 | Both-spoofed frame | 欺骗 | 欺骗 | 两组件均被欺骗 |
数据结构
解压后数据集包含两个主目录:
- Mixed_soundsV1.2: 混合语音+环境音音频,含train/eval划分,每个划分包含wav音频、protocol.txt协议文件和NumPy格式的标签文件夹(含speech、env、mix、original_vs_others_seglab子目录),以及PartialSpoof格式的文本标签
- Partialspoof_background: 源环境音数据,含train/eval划分,包含wav音频、protocol.txt协议文件和标签文件
元数据
每个混合样本关联一个协议条目,记录混合音频ID、语音源、环境音源、混合时长、信噪比、增益/缩放信息和偏移信息。
许可证说明
数据集仅限非商用研究用途。包含多个上游数据集和模型的许可证(CC BY 4.0、CC BY-NC 4.0、CC BY-SA 3.0、CC BY-NC-SA 4.0、CC BY-ND 3.0 US等),用户须遵守所有上游许可条款。其中SBCSAE衍生音频因"禁止修改"条款不直接分发,仅提供重建脚本。




