VISH-DARIJA-TTS: A synthetic moroccan dialect text-audio dataset for vishing and voice-based social engineering...

VISH-DARIJA-TTS v1.0 is a publicly available synthetic multimodal dataset for vishing research in the Moroccan dialect (Darija), a low-resource Arabic variety where publicly documented labelled voice-fraud resources remain highly limited. The dataset was generated through a controlled pipeline comprising scenario authoring, dual-script transcription in Latin-script and Arabic-script Darija, text-to-speech synthesis with Google GenAI Gemini 3.1 Flash TTS, audio normalization, metadata enrichment, deterministic noise augmentation, and release-level validation. It contains 3400 balanced multi-turn scenarios - 1700 normal and 1700 scam/vishing dialogues - with 17,351 dialogue turns. The audio layer includes 3400 clean normalized mono WAV files at 24 kHz and PCM 16-bit, together with 10,200 noisy variants generated at signal-to-noise ratios of 20 dB, 10 dB, and 5 dB, for a total of 13,600 audio files representing approximately 22.61 h of clean audio and approximately 90.45 h when noisy variants are included. The repository includes clean and sanitized text layers, scenario-level and turn-level metadata, heuristic social-engineering labels, emotion and attack taxonomies, audio manifests, train/validation/test splits, SHA-256 checksums, documentation files, and reproducibility scripts. A public text layer replaces sensitive spans with structured placeholders to prevent redistribution of operational social-engineering templates. The dataset contains no real phone calls, real victims, or human-recorded speech. It can be reused for text-based, audio-based, and multimodal vishing detection research, Moroccan Darija speech processing, and controlled noise-robustness benchmarking.