# amine-maazizi/vc-detection-10k (binary_balanced)

## Dataset Purpose
This dataset is built for student-scale anti-spoofing and deepfake audio detection experiments focused on voice conversion (VC).

## Source Corpus
- Real speech source: VCTK subset extracted from local tar shards.

## Fake Generators
- DDDM-VC
- Diff-HierVC

## Label Schema
- Variant: binary_balanced
  - real
  - fake
- Variant: multiclass_by_generator
  - real
  - fake_dddm_vc
  - fake_diff_hiervc

## Split Methodology
Speaker-disjoint train/val/test splits are enforced at subset construction time.

## Limitations
- Generated from two VC systems only.
- Objective labels are pipeline-derived and may include conversion failures if not filtered.
- Short clips (roughly 1-5s) and 16 kHz mono assumptions.

## Storage Assumptions
- WAV audio, 16 kHz, mono, int16.
- Metadata includes duration and sample rate.

## Intended Use
Academic prototyping and benchmarking for VC-focused spoof detection tasks.