Dusha

Dusha数据集是由俄罗斯的Sber机构创建，专注于语音情感识别任务，包含约350小时的音频数据，超过30万条俄语语音及其转录文本。该数据集通过众包平台进行标注，分为表演和真实生活两个子集，前者用于模型预训练，后者用于微调和验证。Dusha数据集旨在通过其丰富的情感表达和多样的语音来源，解决现有数据集在情感识别上的局限性，特别是在面对新说话者时的表现。

The Dusha dataset was developed by Sber, a Russian institution, and focuses on the speech emotion recognition task. It contains approximately 350 hours of audio data, along with over 300,000 Russian speech utterances and their corresponding transcriptions. Annotated via crowdsourcing platforms, the dataset is split into two subsets: the acted subset and the real-life subset. The acted subset is designed for model pre-training, while the real-life subset is reserved for fine-tuning and validation. The Dusha dataset aims to address the limitations of existing speech emotion recognition datasets, especially their performance when dealing with unseen speakers, by providing rich emotional expressions and diverse speech sources.