niloycste68/Bangali_local_dialect_ASR_HF_Dataset

BanglaMix是一个用于方言孟加拉语-英语代码转换自动语音识别（ASR）的大规模数据集，包含41,499个音频片段，总计131.23小时未增强音频，覆盖15种孟加拉国地区方言，并支持英语代码转换。该数据集首次结合了方言和代码转换特性，填补了现有资源仅覆盖单一方面的空白。数据通过YouTube等平台收集，基于方言名称查询和喜剧内容策略，确保语音的自然性和真实性。数据集采用银标准转录方法，使用Whisper large-v3自动生成高置信度转录文本，并包含训练集、开发集和测试集的分割，支持ASR模型的训练和评估。

BanglaMix is a large-scale dataset and benchmark for dialectal Bengali–English code-switching automatic speech recognition (ASR), comprising 41,499 clips with 131.23 hours of unaugmented audio, covering 15 regional dialects of Bangladesh and incorporating English code-switching. It is the first dataset to combine both dialectal and code-switching aspects, addressing the gap in existing resources that cover only one of these features. Data is collected from platforms like YouTube using dialect-name queries and comedy content strategies to ensure natural and authentic speech. The dataset employs silver-standard transcription via Whisper large-v3 for high-confidence automatic transcripts and includes splits for training, development, and test sets, supporting ASR model training and evaluation.