Building Arabic Speech Recognition System Using HuBERT Model and Studying the Sources of Errors

This paper presents the development of a speech recognition system for the Arabic language that can handle continuous speech and a large number of words, independent of the speaker, using deep neural network models trained by self-supervised learning. The system was built using the HuBERT model, and resulted in a word error rate (WER) of 19.3%. Our study on different data sets revealed that the HuBERT-based system has a significant ability to generalize to different spoken dialects. Additionally, we conducted a statistical analysis on the errors specific to the Arabic language that arise from the HuBERT-based system, which highlighted the necessity of incorporating an error correction language model to enhance system accuracy. After the addition of an Arabic language model, the WER decreased to 10.7%. Overall, this study emphasizes the potential of self-supervised learning-based speech recognition systems for the Arabic language and highlights the importance of incorporating language models to enhance system accuracy.