HadTTS: A Novel Fully Diacritized Arabic Synthesized Speech Dataset from Hadith Texts for Fake Audio Detection

The rapid proliferation of neural Text-to-Speech (TTS) technologies has precipitated a critical need for robust forensic countermeasures against malicious deepfakes. However, the development of detection systems for Arabic is currently impeded by the scarcity of high-fidelity, fully diacritized training corpora. To address this gap, this paper introduces HadTTS, a novel synthetic speech dataset derived from the formal register of Hadith literature. Comprising approximately 17 hours of audio generated by 32 distinct voices using the Microsoft EdgeTTS engine, HadTTS uniquely combines classical linguistic rigor with state-of-the-art neural acoustic modeling. Unlike existing generalpurpose datasets, HadTTS leverages the strict orthography of religious texts to ensure maximum pronunciation accuracy, providing a controlled "gold-standard" benchmark for isolating the spectral and prosodic artifacts of modern neural vocoders. We detail the dataset creation methodology, statistical properties, and its primary utility in training and evaluating the next generation of Arabic fake audio detectors.