VoiceGuard
Deepfake Audio Detection System
Accuracy
AUC
Models in Ensemble
Overview
VoiceGuard is a deepfake audio detection project I built to identify whether a speech recording is real or AI-generated. As voice cloning tools get better, synthetic speech is becoming harder to catch and easier to misuse. To improve detection, I used four different audio representations (spectrograms) instead of relying on just one, then combined predictions from five models in an ensemble. This setup reached 97% accuracy and 97% AUC.
Why Four Spectrograms?
Traditional deepfake detection systems analyze audio through a single spectrogram — capturing only a limited view of the signal. VoiceGuard uses four simultaneously, each revealing different patterns and artifacts that synthetic speech leaves behind.
MFCC
Spectral envelope (timbre) & vocal tract characteristicsCaptures the overall shape of the speech spectrum related to the vocal tract. Useful for spotting unnatural timbral patterns or over-smoothed characteristics that can appear in synthetic or cloned voices.
System Architecture
Deployment Architecture
Chosen for this project - predictable cost, full control, ~$0.07/hr on t3.medium
Beyond English: Arabic Deepfake Detection
Extending VoiceGuard to detect Arabic deepfakes - building a first-of-its-kind dataset.
Collect Arabic Audio
Gather non-copyright Arabic speech recordings from public sources
Record Volunteer Speech
Record clean speech from volunteers with Multiple Arabic dialects, starting from Saudi dialects.
Fine-tune TTS Models
Fine-tune 4 models: F1-TTS, XTTS, FishAudio, Qwen3
Generate Synthetic Audio
Generate Arabic deepfake audio, starting with Saudi dialects and expanding to other dialects
Train Arabic Detector
Train detection model on combined real + synthetic Arabic dataset
The dataset is currently being expanded to include dialects from 11 Arab countries.
Results
F1 Score
97%
AUC Score
97%
Inference Time
< 30s per file
Benchmark
Outperforms several peer-reviewed papers
Robustness
Works across voices & recording quality
Papers I Outperform
Research paper in progress.