SONAR
收藏资源简介:
SONAR数据集由福特汉姆大学和IBM研究团队创建,旨在评估和区分最先进的AI合成音频内容。该数据集包含来自9个不同音频合成平台的合成语音,涵盖了领先的TTS服务提供商和最新的TTS模型。数据集的创建过程包括使用多个先进的TTS服务提供商和开源API生成合成语音,并从多个TTS模型中收集合成音频。SONAR数据集主要应用于AI合成音频检测领域,旨在解决AI合成音频的识别和防范问题,特别是在防止恶意用途如深度伪造、欺诈和传播虚假信息方面。
The SONAR dataset was developed by researchers from Fordham University and IBM Research, with the core objective of evaluating and differentiating state-of-the-art AI-synthesized audio content. This dataset contains synthesized speech from nine distinct audio synthesis platforms, covering leading text-to-speech (TTS) service providers and cutting-edge TTS models. The dataset construction process involves generating synthesized speech using multiple leading TTS service providers and open-source APIs, as well as collecting synthesized audio from various TTS models. The SONAR dataset is primarily applied in the field of AI-synthesized audio detection, aiming to address the issues of identifying and guarding against AI-generated audio, particularly to prevent malicious uses such as deepfakes, fraud, and the spread of disinformation.
SONAR: A Synthetic AI-Audio Detection Framework and Benchmark
数据集
该数据集包括以下几个子数据集:
数据集的目录结构应如下所示:
data ├── LJSpeech-1.1 │ ├── wavs │ ├── metadata.csv │ └── README ├── wavefake │ ├── ljspeech_full_band_melgan │ ├── ljspeech_hifiGAN │ ├── ... │ └── ljspeech_waveglow ├── LibriSeVoc │ ├── diffwave │ ├── gt │ ├── ... │ └── wavernn ├── in_the_wild │ ├── 0.wav │ ├── ... │ ├── 31778.wav │ └── meta.csv
使用示例
训练传统模型
运行 main_tm.py 文件进行训练。
-
参数:
--config:不同模型的配置文件。
-
训练 AASIST 模型:
python main_tm.py --config ./config/AASIST.conf
-
评估模型(修改配置文件中的
model_path):python main_tm.py --config ./config/AASIST.conf --eval
微调基础模型
运行 main_fm.py 文件进行微调。
-
微调 Wave2Vec2BERT 模型:
python main_fm.py --model wave2vec2bert
致谢
该项目基于以下开源项目构建:

- 1SONAR: A Synthetic AI-Audio Detection Framework~and Benchmark福特汉姆大学 · 2024年



