IndoDialect A Multi-Dialect Conversational Dataset for Indonesian Local Languages

Dialectal variation is a defining characteristic of many low-resource languages, yet it is often simplified in NLP pipelines, which typically assume a single standardized form. While prior work has explored dialect identification and normalization across multiple languages, Indonesian local languages remain under-resourced at the dialect level, particularly in conversational settings. In this thesis, we introduce IndoDialect, a curated conversational dataset comprising 14.2K dialogues across four Indonesian local languages and eleven dialects, including Javanese, Sundanese, Minangkabau, and Dayak varieties. The dataset consists of two components: (i) Talkshow Data, which provides parallel dialogue across dialects, and (ii) Daily Topic Conversation (DTC), which captures informal and natural dialect usage. We benchmark this dataset on two tasks: (i) dialect identification and (ii) dialect-level machine translation. Our results show that fine-tuned encoder-based models and traditional classifiers achieve strong performance for dialect identification, while few-shot generative models perform substantially worse. For machine translation, consistent with prior work on closely related language varieties, phrase-based statistical methods outperform neural approaches under limited data conditions. IndoDialect provides a new resource for studying dialectal variation in Indonesian local languages and establishes a benchmark for evaluating models in low-resource, closely related language settings