regak/Swahili-Deepfake-dataset

Creation of Real and fake Swahili voices # Swahili-Deepfake-dataset A multi-corpus, multi-generator Swahili audio deepfake detection dataset and construction pipeline. Rather than recording a new corpus from scratch, this project aggregates existing public Swahili speech corpora (e.g. Mozilla Common Voice, FLEURS), selects a phoneme-balanced and speaker-diverse subset, generates synthetic ("deepfake") counterparts with multiple TTS/voice-cloning systems, and produces a labeled real/fake dataset with speaker-independent splits for training and benchmarking detection models. See `docs/METHODOLOGY.md` for the full methodology and `docs/PAPER_OUTLINE.md` for the paper this dataset supports. ## Repository structure ``` swahili_deepfake_dataset/ Core pipeline logic (no heavy ML deps required) phonemes.py Swahili grapheme-phoneme tokenizer + frequency counting selection.py Greedy phoneme-coverage + speaker-diversity selection manifest.py Real/fake labeling + speaker-independent train/val/test splits audio_preprocess.py Audio standardization (16 kHz mono WAV) and quality filtering deepfake_gen.py Pluggable interface for TTS/voice-cloning backends scripts/ CLI entry points select_subset.py Corpus metadata -> phoneme-balanced subset extract_selected_clips.py Corpus archive + selected subset -> just those clips (skips full extraction) preprocess_audio.py Raw audio -> standardized, filtered WAV generate_deepfakes.py Selected subset -> synthetic speech (per generator) build_manifest.py Real + fake -> final labeled dataset manifest sync_drive.py Sync data/ to/from Google Drive (Colab session persistence) tests/ Unit tests for the core pipeline logic docs/ Methodology and paper outline notebooks/ Kaggle/Jupyter notebooks (e.g. corpus download) data/ Not committed to git (see .gitignore); lo …