RaphaelOlivier/librispeech_asr_adversarial
收藏Hugging Face2022-08-03 更新2024-03-04 收录
下载链接:
https://hf-mirror.com/datasets/RaphaelOlivier/librispeech_asr_adversarial
下载链接
链接失效反馈官方服务:
资源简介:
该数据集是LibriSpeech的一个子集,经过对抗性修改,旨在欺骗自动语音识别(ASR)模型预测目标输出而非正确输出。数据集包含多个分割,每个分割由相同的语音样本组成,但使用了不同类型和量的噪声进行修改。噪声类型包括半径为0.04的对抗性噪声、半径为0.015的对抗性噪声,以及半径为0.015的对抗性噪声与房间脉冲响应(RIR)噪声的组合。此外,还提供了原始输入(`natural`分割)。每个分割包含两个文本键:`true_text`表示原始LibriSpeech标签,即实际听到的句子;`target_text`表示对抗性攻击的目标句子。一个被该数据集欺骗的ASR模型在`target_text`上会获得较低的词错误率(WER),而在`true_text`上会获得较高的WER。一个对该数据集具有鲁棒性的ASR模型则会相反。
This dataset is a subset of LibriSpeech that has been adversarially modified to deceive automatic speech recognition (ASR) models into predicting target outputs instead of the correct ones. It comprises multiple splits, each consisting of the same speech samples but modified with different types and amounts of noise. The noise types include adversarial noise with a radius of 0.04, adversarial noise with a radius of 0.015, and a combination of adversarial noise with a radius of 0.015 and room impulse response (RIR) noise. Additionally, the original input (the `natural` split) is provided. Each split contains two text keys: `true_text` refers to the original LibriSpeech label, i.e., the actual spoken sentence; `target_text` refers to the target sentence for the adversarial attack. An ASR model fooled by this dataset will achieve a low word error rate (WER) on `target_text` but a high WER on `true_text`. Conversely, an ASR model robust to this dataset will exhibit the opposite performance.
提供机构:
RaphaelOlivier
原始信息汇总
数据集概述
数据集描述
本数据集是LibriSpeech的一个对抗性修改子集,旨在诱导ASR模型预测特定目标而非正确输出。
数据集分割
数据集包含多个分割,每个分割包含相同的语音,但添加了不同类型和量的噪声。具体分割包括:
- 对抗性噪声半径0.04 (
adv_0.04分割) - 对抗性噪声半径0.015 (
adv_0.015分割) - 对抗性噪声半径0.015结合房间脉冲响应(RIR)噪声 (
adv_0.015_RIR分割) - 原始输入 (
natural分割)
数据集内容
每个分割提供两个文本键:
true_text:原始LibriSpeech标签,即实际听到的句子。target_text:对抗攻击的目标句子。
评估指标
数据集通过WER(词错误率)评估ASR模型。一个被此数据集愚弄的模型在target_text上WER较低,在true_text上WER较高;而对此数据集具有鲁棒性的模型则相反。
示例WER结果
| 分割 | target_text WER | true_text WER |
|---|---|---|
| 0.015 | 58.2 | 108 |
| 0.04 | 49.5 | 108 |



