IndicTalk: A Large-Scale Persona-Based Multilingual Conversational Corpus for Indic Languages

Researchers released IndicTalk, a multilingual code-mixed conversational corpus with over 1.3 million multi-turn dialogues across 18 varieties in 9 Indic languages.

The dataset was fully automated, using news grounding, persona-based generation, multilingual LLMs, and quality checks.

It provides a large validated resource for training and evaluating conversational AI in underrepresented Indic languages.

IndicTalk is especially useful for natural code-mixed speech and is available via Hugging Face.