gereon/voxguard-synthetic-speech

VoxGuard Synthetic Speech Dataset

A dataset of 10,000+ AI-generated (deepfake) speech samples created for training and evaluating deepfake speech detection models.

Dataset Description

Samples	~10,040 synthetic WAV files
Format	WAV, 16kHz mono
Generation	Voice cloning via Qwen3-TTS (Replicate API)
Source speakers	280+ unique speakers from LibriSpeech
Source subsets	clean (train.100, train.360, validation, test) + other (train.500, validation, test)

Generation Process

1. Downloaded ~10,040 real speech samples from LibriSpeech (2-30s duration, diverse speakers)
2. Transcribed each sample using OpenAI Whisper (medium)
3. Generated voice clones using Qwen3-TTS in \ mode:
   • Reference audio: original LibriSpeech sample
   • Reference text: Whisper transcription
   • Target text: one of 100 diverse sentences (news, conversation, instructions, etc.)
4. Result: synthetic speech that mimics each speaker's voice saying a different sentence

Contents

• \ - ~10,040 synthetic WAV files
  • \ to \ (original batch, 4-digit naming)
  • \ to \ (extended batch, 5-digit naming)
• \ - Whisper transcriptions for each source sample

Usage

This dataset is used to train the VoxGuard deepfake detection LoRA adapter, which fine-tunes the DF Arena 1B base model.

Note: Only synthetic (fake) speech is included. The corresponding real speech samples are from LibriSpeech and should be obtained separately.

Related

• Model: gereon/voxguard-lora - LoRA adapter for deepfake detection
• Base model: Speech-Arena-2025/DF_Arena_1B_V_1
• Code: gereonelvers/voxguard