Building a neural speech recognizer for quranic recitations

This work is an effort towards building Neural Speech Recognizers system for Quranic recitations that can be effectively used by anyone regardless of their gender and age. Despite having a lot available online, most them are recorded professional male adult reciters, which means ASR trained on such datasets would not female/child reciters. We address this gap adopting benchmark dataset audio records consists both genders from different ages. Using dataset, we build several speaker-independent NSR systems based the DeepSpeech model use word error rate (WER) evaluating them. The goal to show how tuned certain perform test set other gender. Unfortunately, number female in our rather small while much larger. In first experiments, avoid imbalance issue between two down-sample part match part. For subset results interesting with 0.968 WER when tested recitations. same gives 0.406 On hand, training testing it recitation 0.966 0.608 WER.