LibriFake
收藏资源简介:
LibriFake是一个专为评估少样本和单样本语音深度伪造音频检测方法而设计的大规模数据集。它基于LibriSpeech数据集,通过使用最先进的语音克隆模型生成每个样本的对应合成版本。数据集被分为训练和测试子集,测试集包含约15%的说话者,以确保模型对新说话者的泛化能力。LibriFake旨在帮助研究人员评估少样本学习方法和单样本学习方法的性能,特别是在面对新型TTS模型时的适应性和鲁棒性。
LibriFake is a large-scale dataset specifically developed for evaluating few-shot and one-shot speech deepfake audio detection approaches. It is constructed based on the LibriSpeech dataset, with synthetic counterparts for each sample generated via state-of-the-art speech cloning models. The dataset is partitioned into training and test subsets, where the test set includes approximately 15% of the speakers, aiming to validate the model's generalization capability to unseen speakers. LibriFake is intended to help researchers assess the performance of few-shot learning and one-shot learning methods, especially their adaptability and robustness against novel TTS models.
ADD-GP数据集概述
数据集基本信息
- 官方实现:Few-Shot Speech Deepfake Detection Adaptation with Gaussian Processes
- 相关论文链接:未提供
- LibriFake链接:未提供
数据集内容
- 训练数据路径:
/data/Libri_dataset_11labs_new_train.csv - 验证数据路径:
/data/Libri_dataset_11labs_new_val.csv - 测试数据路径:
/data/Libri_dataset_11labs_new_test.csv
技术参数
- 训练批次大小:70
- 验证批次大小:230
- 训练TTS系统:f5, yourtts, valle, Tacotron2, whisper
- 验证TTS系统:eleven_labs
- 支持few-shot学习:100个样本
- 预训练模型检查点路径:
/path/to/pretrained/xlsr/xlsr4_model_epoch_0.pt
运行要求
- Python版本:3.9
- 依赖安装:
pip install -r requirements.txt

- 1Few-Shot Speech Deepfake Detection Adaptation with Gaussian Processes巴伊兰大学,以色列 · 2025年



