TL;DR
Researchers have discovered that the spectral magnitude of human speech, when averaged across many speakers, follows Benford's Law. This observation has led to the development of novel "Benford Similarity" (BenS) features that can effectively distinguish between human and synthetic speech, achieving over 90% accuracy.
What Problem Does This Paper Address?
Benford's Law, a phenomenon where the leading digits of numbers in many naturally occurring datasets follow a predictable logarithmic distribution, has been observed in diverse fields from financial data to image compression. However, its applicability to human speech has remained unexplored. This paper investigates whether human speech spectra adhere to Benford's Law and, if so, leverages this property to create new features for detecting synthetic speech (deepfakes), a growing concern in the digital age.
Methods
The core of the methodology involves analyzing the magnitude spectrum of speech signals. For each frame of a speech segment, the short-time Fourier transform (STFT) is computed. The DC component of the signal is removed, and the magnitude spectrum is scaled to ensure all values are positive. The leading digits of these magnitudes are then extracted to form an empirical probability distribution. This empirical distribution is compared to the ideal Benford distribution using the Kullback-Leibler (KL) divergence, a measure of how one probability distribution differs from another.
To assess Benford's Law adherence for human speech, the KL divergence values are averaged across all speakers. For feature extraction, a set of 11 statistics (mean, standard deviation, and percentiles) are computed from the KL divergence values across all frames of a speech sample. These statistics form the "Benford Similarity" (BenS) features.
Datasets
- TIMIT Acoustic-Phonetic Continuous Speech Corpus: This dataset was used for the primary analysis of whether human speech follows Benford's Law and for training the BenS feature extraction pipeline. It contains audio recordings from 326 male and 136 female speakers.
- Google Text-to-Speech (TTS) API voices: A set of 20 different US-English voices (male and female) from Google's TTS API were used to generate synthetic speech samples for classification experiments.
Key Insights and Ideas
- Speech Spectra Follow Benford's Law (on average): The paper provides empirical evidence that the leading digits of human speech magnitude spectra, when averaged across many speakers, closely follow Benford's Law. This suggests an underlying mathematical property in speech production.
- Benford's Law as a Detector for Synthetic Speech: The BenS features, derived from the deviation from Benford's Law, prove to be effective in distinguishing between human and synthetic speech. This indicates that synthetic speech generation processes may not inherently replicate the Benford-like properties found in natural speech.
- Variability is Key: While the average speech spectrum adheres to Benford's Law, individual speech samples exhibit variability. This variability, captured by the BenS features, is crucial for differentiating between human and synthetic speech.
Key Results
- Benford's Law Adherence: A linear regression analysis comparing the empirical and ideal Benford distributions for human speech yielded an equation with a slope near 1 and an intercept near 0, with an R² value of 0.99, strongly supporting the hypothesis.
- Synthetic Speech Detection Accuracy: A quadratic Support Vector Machine (QSVM) classifier using the BenS features achieved an accuracy of 91.5% in distinguishing between human and synthetic speech, with an error rate of 8.5%.
- Feature Differences: Z-scores for the BenS features showed that synthetic speech samples had significantly higher deviations from the mean compared to human speech samples, indicating a clear difference in their Benford similarity.
Limitations and Future Work
- Sample Size for Classification: The classification experiment was limited by the availability of synthetic voices, resulting in a relatively small dataset of 200 samples for each class. The authors note that the human speech samples used for testing were from a single dialect region (DR2), which might explain a slight positive bias in the BenS features for human speech compared to the broader TIMIT training set.
- Synthetic Speech Generation Models: The study used TTS voices from Google's API. Future work could explore a wider range of synthetic speech generation techniques, including more sophisticated deepfake audio.
- Clinical Applications: The authors suggest that the BenS features could be valuable for detecting early changes in speech due to neurological conditions where the independence assumption in speech production models is violated. This would involve evaluating the utility of BenS features for persistent on-device tracking in clinical settings.
- Deeper Understanding of Production Mechanisms: Further research is needed to fully understand why synthetic speech generation processes differ in their adherence to Benford's Law, potentially by analyzing the underlying complex non-linear structures of neural network-based synthesizers.
Summary updated