DailyTalk

DailyTalk是由韩国科学技术院与KRAFTON公司合作创建的高质量对话式语音数据集，专为对话式文本到语音转换（TTS）设计。该数据集包含2,541个对话，总时长约20小时，源自DailyDialog数据集，经过筛选、修改和录音处理，确保音频质量及对话的自然性。创建过程中，特别注意保持对话的多样性和学术开放性，同时添加了填充词如'uh'和'umm'以增强真实感。DailyTalk的应用领域主要集中在提升TTS系统在对话情境中的表现，解决传统TTS模型在处理连续对话时缺乏上下文连贯性的问题。

DailyTalk is a high-quality conversational speech dataset co-developed by the Korea Advanced Institute of Science and Technology (KAIST) and KRAFTON, tailored explicitly for conversational text-to-speech (TTS) applications. Comprising 2,541 conversations with an aggregate duration of roughly 20 hours, the dataset is derived from the DailyDialog dataset, and has undergone rigorous filtering, content revision, and audio recording workflows to guarantee superior audio quality and natural conversational authenticity. During the development process, special attention was paid to preserving conversational diversity and academic openness, while incorporating filler words such as "uh" and "umm" to bolster the realism of the dialogues. The core application scenarios of DailyTalk center on enhancing the performance of TTS systems in conversational contexts, addressing the limitation of traditional TTS models that lack contextual coherence when processing continuous multi-turn conversations.