Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 10·5 min read
Linguistically Augmented Audio Speech Data (LinguAS)
TL;DR
As deepfaked and spoofed audio becomes more prevalent, existing detection models often overlook important linguistic features. Linguistically Augmented Audio Speech Data (LinguAS) was created, comprising over 800 audio samples annotated with Expert-Defined Linguistic Features (EDLFs).
✦ Why It Matters
Engineers can leverage LinguAS to improve audio deepfake detection models by incorporating linguistic features.
Key Takeaways
- Utilize LinguAS for training audio detection models.
- Incorporate linguistic features to enhance model performance.
- Explore the dataset's metadata for improved analysis.
- Stay updated on advancements in deepfake detection techniques.
Full Summary
The rise of maliciously created fake speech, such as deepfakes, poses a significant challenge for audio detection systems, which typically rely on frame-level audio features. To address this, Linguistically Augmented Audio Speech Data (LinguAS) was developed, featuring over 800 audio samples that include both genuine and deepfaked speech.
Each sample is annotated with five Expert-Defined Linguistic Features (EDLFs) that reflect natural human speech patterns. The dataset also includes metadata on speaker gender and the source of spoofed audio, providing additional context for model training.
Experiments demonstrated that models trained with LinguAS outperformed existing deep learning baselines, such as ASVspoof 2021 and SSL models like HuBert and XLSR. This indicates that incorporating linguistic features can significantly enhance the detection of audio deepfakes.
The dataset and accompanying code are publicly available for further research.