RobotsMali/transcription-scorer

Transcription Scorer数据集是一个为了支持无参考评价自动语音识别系统而创建的人类反馈数据集。该数据集包含2153个音频样本，每个样本都附带有一个转录文本、一个0到100之间的人工标注分数以及标注者的信息。这些音频样本来源于不同的渠道，包括演讲和带有歌词的音乐。数据集可用于开发无参考评价指标、训练基于人类反馈的强化学习模型的ASR系统，以及研究人类偏好与转录质量之间的关系。

The Transcription Scorer dataset is created to support research in reference-free evaluation of Automatic Speech Recognition (ASR) systems using human feedback. It contains 2153 audio samples, each with an associated transcription, a human-annotated score between 0 and 100, and the identity of the labeler. The audio samples are sourced from various channels, including speeches and music with lyrics. The dataset is useful for developing reference-free evaluation metrics, training reward models for ASR systems fine-tuned with human feedback, and studying the relationship between human preferences and transcription quality.