OOD-Speech

OOD-Speech是首个针对孟加拉语自动语音识别（ASR）的分布外（OOD）基准测试数据集。该数据集由Bengali.AI创建，涵盖了来自南亚22,645名母语为孟加拉语的说话者的语音数据，总训练时长为1177.94小时，测试集包含23.03小时的数据，来源于17个不同的语音源，如电视戏剧、有声书、脱口秀等。数据集通过大规模在线众包活动收集，旨在解决孟加拉语方言和韵律特征多样性带来的ASR模型鲁棒性问题。OOD-Speech不仅是最公开可用的孟加拉语语音数据集，也是首个针对孟加拉语的OOD ASR基准测试数据集，适用于评估ASR模型在不同语音分布下的性能。

OOD-Speech is the first out-of-distribution (OOD) benchmark dataset for Bengali automatic speech recognition (ASR). Developed by Bengali.AI, this dataset encompasses speech data from 22,645 native Bengali speakers across South Asia, with a total training duration of 1177.94 hours and a test set of 23.03 hours. The test set is sourced from 17 distinct speech sources such as television dramas, audiobooks, talk shows and more. Collected via large-scale online crowdsourcing campaigns, this dataset aims to address the robustness challenges of ASR models stemming from the diversity of Bengali dialects and prosodic features. OOD-Speech is not only the most publicly available Bengali speech dataset, but also the first OOD ASR benchmark dataset tailored for the Bengali language, which is suitable for evaluating the performance of ASR models under different speech distributions.