PromptDialog

This page presents the results obtained by fine-tuning SoulX-Podcast on our dataset. We compare our model with other models that possess related capabilities across multiple tasks, in order to demonstrate the superiority of our dataset. Since our dataset is primarily composed of Chinese data, we mainly present the performance of our model on Chinese speech generation. In addition, we supplement the training data with 100 hours of English audiobook data for fine-tuning, thereby demonstrating that the construction of our dataset exhibits cross-lingual transferability.