Disentangled Representations for Arabic Dialect Identification based on Supervised Clustering with Triplet Loss
Zainab Alhakeem, Yoohwan Kwon, Hong-Goo Kang · 2021 29th European Signal Processing Conference (EUSIPCO) · 2021
In this paper, we propose a novel supervised clustering with triplet (SCT) loss that effectively learns disentangled representations for Arabic dialect identification (ADI). To improve the performance of ADI using latent representation based approaches, we need to extract embeddings that include only dialect related information by dissociating all the irrelevant information such as gender, channel, and speaker. In consideration of the embedding-level distribution, our proposed SCT loss minimizes intra-class variations and maximizes inter-class variations. Specifically, it uses the centroid of each dialect as a triplet component, thereby avoiding the issue of choosing an undesirable triplet component due to random sampling. Experimental results on the ADI-17 dataset show that our proposed method significantly outperforms conventional state-of-the-art methods in terms of the identification accuracy.