Development and Analysis of a Versatile Dataset of Speech, Real and Synthesized, of Arabic Learners
Mansour Alsulaiman, Mohammed Algabri, Hassan I. Mathkour, Mohamed Abdelkader Bencherif, Ghulam Muhammad, Saad Al‐Gahtani, Mohammed Faisal, Mohamed Amine Mekhtiche, Taha M. Alfakih · 2023
Computer-Aided Pronunciation Training (CAPT) systems are gaining popularity recently due to the advancements in deep neural networks (DNN) and machine learning and the availability of databases of speech of language learners. Unfortunately, research in Arabic CAPT systems suffer from the lack of CAPT datasets compared to other languages. In this paper, we present the details of and the ideas used in the development of a versatile dataset of speech, real and synthesized, of Arabic Learners. To develop the dataset, we utilized an existing Arabic speech corpus, King Saud University Speech Database (KSU-SD). KSU-SD's main application was a speaker recognition system, but it was designed to be useful in other applications such as CAPT systems. KSU-SD includes a large number of speakers from diverse nationalities (Saudis, Arabs, and Non-Arabs). KSU-SD contains the recording of about 60 non-native speakers from more than 20 nationalities, hence we selected it to build the non-native Arabic-CAPT corpus. The developed corpus consists of transcribed, segmented, and annotated speech, which makes it suitable for building Arabic CAPT systems. In addition to presenting the details of developing the dataset, we also present an analysis of the text and speech errors of the dataset. The dataset was verified by many CAPT systems.