sleeping-ai/hebrew-exp

我们提出了一个使用Whisper Turbo模型生成的高质量希伯来语语音转录数据集。相比之下，HebDB（https://pages.cs.huji.ac.il/adiyoss-lab/HebDB/）依赖于Whisper Large，在转录准确性和鲁棒性方面明显不如Whisper Turbo。此外，HebDB以希伯来语文件名分发音频文件，这一设计选择在现代训练流程和预处理工作流程中引入了不必要的摩擦。我们的数据集专门为研究和大型训练而设计，无需额外的规范化或重构，可立即用于训练各种语音和语言模型。

We present a high-quality Hebrew speech transcription dataset generated using the Whisper Turbo model. In contrast, HebDB (https://pages.cs.huji.ac.il/adiyoss-lab/HebDB/) relies on Whisper Large, which is demonstrably inferior to Whisper Turbo in transcription accuracy and robustness. Furthermore, HebDB distributes audio files with Hebrew filenames, an avoidable design choice that introduces unnecessary friction into modern training pipelines and preprocessing workflows. Our dataset is deliberately engineered for research and large-scale training, requiring no additional normalization or restructuring, and is immediately usable for training a wide range of speech and language models.