Audio Deepfake Detection

Examines the frequency-domain representation of audio to identify anomalies invisible to the human ear. Neural vocoders leave distinct grid-like patterns in high-frequency bands due to upsampling operations. Key techniques include:

• Mel-Frequency Cepstral Coefficients (MFCC) extraction to model human auditory perception
• Detection of unnatural harmonic structures absent in biological vocal tracts
• Analysis of sub-band energy distributions for synthetic discontinuities
• Identification of phase inconsistencies that linear predictive models cannot replicate

Evaluates the suprasegmental features of speech—pitch contour, rhythm, stress patterns, and intonation—for unnatural temporal dynamics. Synthetic speech often exhibits:

• Overly regular pitch trajectories lacking natural micro-variation
• Unnatural pause distributions that violate linguistic prosodic boundaries
• Flat or monotonic intonation curves inconsistent with emotional context
• Duration anomalies where phoneme lengths deviate from human baselines

These features are fed into sequential models trained on genuine speech corpora.

Distinguishes between a live human speaker and a presentation attack—a pre-recorded or synthetic audio sample played back to a sensor. Core techniques include:

• Pop noise detection from plosive consonants (p, b, t) that speakers physically produce
• Channel mismatch analysis between the expected acoustic environment and the injected signal
• Challenge-response protocols requiring spontaneous, unscripted verbal responses
• Multi-modal fusion with video-based lip-sync consistency verification

Standardized under ISO/IEC 30107 for biometric presentation attack detection.

Bypasses traditional feature engineering by feeding raw audio samples directly into end-to-end deep learning architectures. RawNet2 and AASIST models learn to detect:

• Artifacts in the time domain from neural vocoder upsampling layers
• Sub-sample level discontinuities introduced during waveform synthesis
• Unnatural sample correlations that differ from physical acoustic recordings
• Vocoder-specific ringing artifacts in the decoded waveform

This approach avoids information loss from handcrafted feature extraction pipelines.

Evaluates the coherence of acoustic features across consecutive analysis windows to identify frame-level manipulation. Synthetic audio often exhibits:

• Jitter and shimmer anomalies—unnatural cycle-to-cycle variations in fundamental frequency and amplitude
• Discontinuities at vocoder frame boundaries where concatenative artifacts appear
• Non-physical formant transitions that violate the physiological constraints of the human vocal tract
• Statistical divergence in sequential embedding spaces from genuine speech distributions

The application of MFCC features—which model the non-linear frequency perception of the human auditory system—as input to classifiers that distinguish genuine speech from deepfakes. Synthetic vocoders often fail to replicate the natural correlations between adjacent cepstral coefficients, leaving detectable artifacts.

• Extracts short-term power spectrum of sound on the mel scale
• Captures the shape of the vocal tract filter, not just raw frequency
• Deepfake detectors analyze delta and delta-delta coefficients to capture temporal dynamics
• Common classifiers: Gaussian Mixture Models, ResNet-based architectures

The forensic identification of the specific neural vocoder architecture (e.g., WaveNet, HiFi-GAN, WaveGlow) used to synthesize speech from mel-spectrograms. Each vocoder introduces unique, imperceptible artifacts in the high-frequency spectrum and phase components.

• HiFi-GAN leaves characteristic grid patterns in the 8-16 kHz band
• WaveNet exhibits autoregressive error accumulation in long utterances
• Fingerprints persist even after transcoding or compression
• Enables attribution to specific synthesis toolchains

The detection of unnatural patterns in suprasegmental features—pitch contour, rhythm, stress, and intonation—that synthetic speech generators struggle to model authentically. Human speech exhibits micro-variations in prosody tied to emotional state and cognitive load.

• Measures fundamental frequency (F0) jitter and shimmer
• Synthetic speech often shows unnaturally flat or overly smooth pitch contours
• Analyzes pause duration distributions and speaking rate variability
• Deepfakes fail to replicate declination—the gradual pitch drop across an utterance

A forensic technique that examines the phase component of the Fourier transform, which most neural vocoders reconstruct poorly compared to the magnitude spectrum. Human speech contains structured phase relationships that synthetic generators fail to preserve.

• Group delay analysis reveals vocoder artifacts invisible in magnitude spectra
• Synthetic phase often appears randomized or unnaturally uniform
• Modified Group Delay Function (MGDF) highlights phase discontinuities
• Effective against both GAN-based and autoregressive vocoders

The forensic analysis of whether synthetic speech exhibits the Lombard reflex—the involuntary human vocal adaptation to noisy environments involving increased amplitude, pitch, and vowel duration. Current deepfake models fail to generate contextually appropriate Lombard speech.

• Humans automatically adjust vocal effort based on ambient noise
• Synthetic speech maintains constant spectral tilt regardless of simulated environment
• Absence of expected Lombard modifications indicates synthetic origin
• Particularly useful for detecting faked field recordings or emergency calls

A statistical forensic method that compares the duration histograms of individual phonemes against distributions derived from large corpora of natural human speech. TTS systems often exhibit compressed or unnaturally uniform phoneme timing.

• Models Gaussian mixture distributions for each phoneme class
• Synthetic speech shows reduced variance in vowel duration
• Coarticulation timing between adjacent phonemes is a key discriminative feature
• Bi-LSTM networks trained on phoneme-aligned corpora achieve state-of-the-art results