Efficient Romanian Dialect Identification in Low-resource conditions using Transfer-Learning and Metric-Learning

Automatic spoken dialect identification (DID) is a challenging task as dialects of a given language have high similarities. Lack of sufficient training data to train deep learning (DL) based DID systems, which is common for low-resource languages like Romanian, makes it further challenging. While transfer-learning can be used to improve the performance in such low-resource conditions, it is not a trivial task to select a pre-trained model and a suitable layer for DID-specific feature extraction, as the ability to encode DID-specific content vary largely across models and their layers. Furthermore, due to high inter-class similarities, the features obtained from such pre-trained models may contain only limited dialect-discriminability, necessitating explicit methods to improve the discriminability of the learned features. Motivated by all these, we propose different methods to build efficient DID systems for identifying Romanian dialects, in low-resource settings. Specifically, we first explore pre-trained models like wav2vec2, HuBERT and XEUS for Romanian DID along with determining their optimum layer for this task. Furthermore, to address the performance degradation due to high inter-class similarities and intra-class variations, we employ metric learning techniques like center loss (CL) and centroid similarity loss (CSL). Obtained results indicate that DID system trained using 4^th layer of XEUS model, when combined with CSL gives the best performance on Romanian DID.