Deepfake Audio Detection via Phoneme Turbulence Analysis

The patent segments the speech signal into individual phoneme instances and analyzes each phoneme's turbulent flow characteristics independently. By examining multiple phoneme-level measurements rather than treating the entire speech signal as a single entity, the system captures diverse turbulent flow patterns across different phonemes, enabling generalization to unseen deepfake generation techniques while maintaining high detection precision.

The patent replaces machine learning-based discriminators with a physics-based turbulence analysis approach. Instead of using trained models that depend on observed generation techniques, the system measures physical turbulent flow characteristics of speech signals, which are inherent to organic speech production and independent of synthetic generation methods. This substitution enables generalization across datasets while maintaining detection capability.

The patent employs a self-service detection mechanism where the system measures intrinsic turbulent flow characteristics of speech signals without requiring external training data or knowledge of generation techniques. By relying on physical properties that are inherent to organic speech production, the system automatically adapts to new deepfake methods without retraining, maintaining both precision and reliability against adaptive adversaries.

The patent changes the detection parameter from high-level spectral features used by machine learning models to fundamental physical parameters of turbulent flow. By measuring parameters such as frequency response variations and temporal dynamics of turbulent phonemes, the system detects deepfakes based on physical inconsistencies rather than statistical patterns, making it reliable against adaptive adversaries who cannot easily replicate physical turbulence.