---
language: en
license: mit
tags:
  - audio
  - deepfake
  - synthetic-speech
  - tts
  - voice-cloning
  - scam-detection
task_categories:
  - audio-classification
size_categories:
  - 1K<n<10K
---

# VoxGuard Synthetic Speech Dataset

A dataset of AI-generated speech samples for training and evaluating deepfake speech detection models.

## Contents

### 1. Voice-Cloned Speech (`fakes/`)
- **Samples:** 1,040 WAV files
- **Generation:** Voice cloning with [Qwen3-TTS](https://huggingface.co/Qwen/Qwen3-TTS) via Replicate API
- **Source:** [LibriSpeech](http://www.openslr.org/12/) (clean subset, 123 unique speakers)
- **Format:** WAV, 16kHz mono
- **Process:** Original audio → Whisper transcription → Qwen3-TTS voice clone
- **Metadata:** `transcriptions.json`

### 2. Synthetic Scam Calls (`scam_calls/`)
- **Samples:** 100 WAV files
- **Generation:** [ElevenLabs Text-to-Dialogue API](https://elevenlabs.io/docs/api-reference/text-to-dialogue/convert) (Eleven v3 model)
- **Format:** WAV, 16kHz mono
- **Categories:** IRS/tax, tech support, prize/lottery, bank fraud, investment/crypto, romance, charity, insurance/Medicare, job offer, utility/service
- **Obviousness levels:** obvious, moderate, subtle
- **Voices:** 20 distinct ElevenLabs voices across conversations
- **Metadata:** `scam_calls/metadata.json` (includes full transcripts)

## Usage

This dataset was used to train the [VoxGuard LoRA](https://huggingface.co/gereon/voxguard-lora) deepfake detection adapter.

**Note:** Only synthetic (fake) speech is included. Corresponding real speech samples should be obtained separately.

## Related

- Model: [gereon/voxguard-lora](https://huggingface.co/gereon/voxguard-lora)
- Code: [gereonelvers/voxguard](https://github.com/gereonelvers/voxguard)