Researchers from ARTPARK-IISc released Vaani-LID_v0, a language identification model trained on 10 hours each of 42 Indian languages drawn from the roughly 31,255-hour Vaani speech dataset spanning 165 districts. The study compared Whisper and Indic-pretrained FastConformer encoders, testing each frozen and fine-tuned under three training objectives: cross-entropy, supervised contrastive loss, and hierarchical softmax. The Vaani-pretrained FastConformer generalized best to out-of-domain data even while frozen, and hierarchical softmax consistently outperformed flat classification by exploiting linguistic family structure.
