Mobile Voice Cloning dataset

This dataset supports the research paper titled \A Supervised Transformer-based Model for Attributing Mobile-App-generated Synthetic Audio Artifacts\ by Lucky Onyekwelu-Udoka and Yong Guan. The dataset includes labeled audio samples collected from both natural sources and four popular mobile synthesizer applications ClonyAI, VoiceAI, and Voice Changer (in two different modes: Text-to-Audio and Voice Modulation). Each audio sample is annotated with its source class to facilitate supervised learning for synthetic audio detection and attribution.Audio signals were preprocessed into 2D mel spectrograms and augmented using techniques like pitch shifting, time scaling, and amplitude adjustment. The dataset was used to train a Custom PaSST (Patchout Spectrogram Transformer)-based model integrated with a multi-layer perceptron classifier. This model achieved 71% validation accuracy in attributing audio artifacts to their source applications and 80% in binary classification (natural vs. synthetic).The dataset consists of 5 classes:Class 0: Natural audio (Flickr Audio Dataset)Class 1: ClonyAI-generated audioClass 2: VoiceAI-generated audioClass 3: Voice Changer (Text-to-Audio)Class 4: Voice Changer (Voice Modulation)This dataset can aid in future research on audio forensics, synthetic media detection, and AI-based content attribution models.