RETRACTED: Designing an ASR Corpus for the Albanian Language

Abstract

This paper reports on the creation of a corpus for the Albanian language that is intended for training and evaluating Automatic Speech Recognition (ASR) systems. The corpus comprises 100 h of audio recordings taken from 200 audiobooks and covers a wide range of topics with a rich vocabulary. The audio recordings were transcribed manually, strictly verbatim, and listened to carefully several times to ensure accuracy. The corpus was evaluated using various end-to-end models as well as Transformer-based architectures. The evaluation was conducted on both the training and testing sets, with Word Error Rate (WER) and Character Error Rate (CER) being considered as evaluation metrics. The results of the architectures trained with this corpus were compared with the results of the LibriSpeech corpus in English. The best architecture based on end-to-end models yielded 5% WER and 1% CER on the training set and 35% WER and 11% CER on the testing set. The transformer-based architecture yielded great results in the testing set, reaching a WER of 18%.

Keywords

• Albanian language
• automatic speech recognition
• corpus