AudioHate-DB

This dataset contains audio recordings of English speech spoken by Bangladeshi speakers, which can be used for research on hate speech detection, processing of accented-English speech and audio-based natural language understanding. The dataset aims to fill an important gap in the current resources, as it records English hate and non-hate speech as expressed in the phonetic and prosodic features of the Bangladeshi accent. The data set consists of a total of 3018 audio samples, with 1515 samples of hate speech and 1503 samples of non-hate speech. The almost balanced distribution allows to train and test binary classification models without employing elaborate techniques to balance the classes. The data contained within it are divided into two kinds: 1. Hate Speech: 1,515 audio samples with English utterances that are classified as hate speech. 2. Non-Hate Speech: 1,503 audiosamples that include neutral or non-offensive English speech. Both versions of the data are available. Original recordings of audio data in their natural state without changes and processed data: audio recordings that have been cleaned and redacted for any potential noise to provide standardized audio that can be directly used in any pipelines. Audio files are available in WAV format and are sorted into distinctly labeled folders based on the given class (hate / non_hate) and version (raw / processed).本数据集收录了孟加拉国使用者录制的英语语音音频，可应用于仇恨言论检测（hate speech detection）、带口音英语语音（accented-English speech）处理以及基于音频的自然语言理解（audio-based natural language understanding）相关研究。当前相关资源存在重要空白，本数据集旨在填补这一空白——其所收录的英语仇恨与非仇恨言论，均体现了孟加拉口音的语音学与韵律学特征。 本数据集共计包含3018条音频样本，其中仇恨言论样本1515条，非仇恨言论样本1503条。两类样本的分布近乎均衡，无需借助复杂的类别平衡技术即可用于二分类模型（binary classification models）的训练与测试。 本数据集包含两类数据： 1. 仇恨言论：1515条英语语音音频样本，经标注属于仇恨言论范畴。 2. 非仇恨言论：1503条包含中立或无冒犯性英语语音的音频样本。 本数据集同时提供两种版本的音频数据：其一为未经修改的原始自然状态音频录音；其二为经过降噪处理并清除潜在杂音的标准化音频，可直接集成至各类处理流水线（pipeline）中。音频文件均采用WAV格式存储，并按照类别（仇恨/非仇恨）与版本（原始/处理后）分别置于带有明确标签的文件夹中。