Reproduce “J-CHAT: Japanese Large-scale Spoken Dialogue Corpus for Spoken Dialogue Language Modeling” on OpenResearch

About this paper

Wataru Nakata, Kentaro Seki, Hitomi Yanaka, Yuki Saito, Shinnosuke Takamichi, Hiroshi Saruwatari

Researchers from The University of Tokyo present J-CHAT, a 76,036-hour open-source Japanese spoken dialogue corpus, created using an automated, language-independent methodology to gather diverse, spontaneous, and acoustically clean data from internet sources. The corpus demonstrates superior acoustic quality and diversity compared to existing resources, enabling generative spoken dialogue models trained on J-CHAT to achieve higher naturalness (2.28 MOS) and meaningfulness (2.18 MOS) in subjective evaluations.

Key insights

• An automated, language-independent methodology can effectively address the limitations of manual recording, enabling the scalable creation of large, acoustically clean, and diverse spoken dialogue corpora from web-sourced data.
• Dialogue corpora derived from diverse "in-the-wild" internet sources (YouTube, podcasts) exhibit significantly higher linguistic and phonetic diversity than smaller, often scripted, existing datasets, which is crucial for training robust SDSs.
• While increasing dataset size is beneficial, simply adding more data from a single source may not proportionally improve spoken dialogue generative model performance, highlighting the importance of collecting data from multiple, diverse sources.

Summary by alphaXiv.