Using Character-Level Sequence-to-Sequence Model for Word Level Text Generation to Enhance Arabic Speech Recognition

Owing to the linguistic richness of Arabic language, which contains more than 6000 roots, building a reliable language model for speech recognition systems faces many challenges. This paper introduces free automatic system Modern Standard based on an end-to-end-based Deep Speech architecture developed by Mozilla. The proposed uses character-level sequence-to-sequence map character alignment produced recognizer onto corresponding words. outperformed recent studies single-speaker and multi-speaker using two different state-of-the-art datasets. first was Multi-Genre Broadcast (MGB2) corpus with 1200 h audio data from multiple speakers. achieved new milestone in MGB2 challenge word error rate (WER) 3.2, outperforming related work same reduction 17%. An additional experiment 7-hour Saudi Accent Single Speaker Corpus (SASSC) used build single male speaker-based network architecture. experiments WER 4.25 relative improvement 33.8%.