Golos

Golos是一个专为俄语语音研究设计的大型数据集，由俄罗斯的Sber机构创建。该数据集主要包含通过众包平台手动标注的音频文件，总时长约1240小时。数据集内容丰富，涵盖多种语音数据，包括从众包平台和智能屏幕SberPortal收集的音频。创建过程中，通过模板创建、音频生成、众包验证和辅助转录四个步骤确保数据质量。Golos数据集主要用于语音识别系统的训练和测试，旨在提高自动语音识别（ASR）算法的性能和鲁棒性。

Golos is a large-scale dataset specifically designed for Russian speech research, created by Sber, a Russian institution. This dataset primarily consists of manually annotated audio files collected via crowdsourcing platforms, with a total duration of approximately 1,240 hours. The dataset covers rich and diverse speech data, including audio collected from both crowdsourcing platforms and the SberPortal smart screen. During the dataset's creation, four steps including template creation, audio generation, crowdsourcing verification and auxiliary transcription were adopted to ensure data quality. The Golos dataset is mainly used for training and testing speech recognition systems, with the goal of enhancing the performance and robustness of automatic speech recognition (ASR) algorithms.