Parismita Gogoi, Sishir Kalita, Wendy Lalhminghlui, Viyazonuo Terhiija, Moakala Tzudir, Priyankoo Sarmah, S. R. M. Prasanna
The study evaluates SSL models for tone recognition in three low-resource North-East Indian languages, finding that Mizo tones are recognized best and highlighting the importance of middle layers in SSL models.
Researchers examined how well self-supervised learning (SSL) models can recognize tones in three low-resource languages from North-East India: Angami, Ao, and Mizo. They tested four different models to see how they perform with these languages' tones. The study found that the models recognized tones best in the Mizo language and worst in Angami. The analysis also showed that the middle layers of these models are crucial for recognizing tones, regardless of whether the models were initially trained on tonal or non-tonal languages.
