Arabic speech recognition using deep learning and common voice dataset

Abstract

Speech recognition is critical in creating a natural voice interface for human-to-human communication with modern digital life equipment. Smart homes, vehicles, autonomous devices in the Internet of Things, and others need to recognize various spoken languages. Meanwhile, the Arabic language has a shortage of speech recognition systems. This study comes to develop an Arabic speech-to-text tool for Arabic language. Our solution uses DeepSpeech model which is a deep learning approach and uses a data set from the Common Voice Mozilla project. The results showed a 24.3 percent Word Error Rate and a 17.6 percent character error rate. So, the proposed model reduces the Word Error Rate by 11.7% compared to Bakheet's Wav2Vec model