VISH-DARIJA-TTS: A Synthetic Moroccan Darija Text-Audio Dataset for Vishing and Voice-Based Social Engineering Detection

VISH-DARIJA-TTS is a synthetic Moroccan Darija text-audio dataset designed for research on vishing, voice phishing, and voice-based social engineering detection. The dataset contains 3,400 multi-turn scenarios, including 1,700 scam and 1,700 non-scam dialogues, represented in Latin-script Moroccan Darija and Arabic-script Darija. It includes 3,400 normalized clean TTS-generated WAV files and 10,200 controlled noisy variants at 20 dB, 10 dB, and 5 dB SNR. The release also includes scenario-level and turn-level metadata, emotion normalization, social-engineering taxonomies, audio manifests, train/validation/test splits, validation reports, checksums, and documentation.

The dataset is synthetic and does not contain real victim calls or human-recorded scam calls. Clean text files are aligned with the audio files, while public text files provide a sanitized version with sensitive operational details replaced by placeholders.