Kurdish spoken dialect recognition using x-vector speaker embeddings

Arash Amani , Mohammad Mohammadamini (1) , Hadi Veisi (2)

1 LIA - Laboratoire Informatique d'Avignon
2 University of Tehran

This paper presents a dialect recognition system for the Kurdish language using speaker embeddings. Two main goals are followed in this research: first, we investigate the availability of dialect information in speaker embeddings, then this information is used for spoken dialect recognition in the Kurdish language. Second, we introduce a public dataset for Kurdish spoken dialect recognition named Zar. The Zar dataset comprises 16,385 utterances in 49h-36min for five dialects of the Kurdish language (Northern Kurdish, Central Kurdish, Southern Kurdish, Hawrami, and Zazaki). The dialect recognition is done with x-vector speaker embedding which is trained for speaker recognition using Vox-celeb1 and Voxceleb2 datasets. After that, the extracted x-vectors are used to train support vector machine (SVM) and decision tree classifiers for dialect recognition. The results are compared with an i-vector system that is trained specifically for Kurdish spoken dialect recognition. In both systems (i-vector and x-vector), the SVM classifier with 86% of precision results in better performance. Our results show that the information preserved in the speaker embeddings can be used for automatic dialect recognition.

Mots clés: Speaker embeddings; x-vector; Kurdish language; dialect recognition; Zar dataset; Speaker embeddings

Domaines: Intelligence artificielle [cs.AI]; Traitement du signal et de l'image [eess.SP]

Arash Amani, Mohammad Mohammadamini, Hadi Veisi. Kurdish spoken dialect recognition using x-vector speaker embeddings. 2021. ⟨hal-03262435⟩