Cepstral-Domain Metrics in Signal Analysis

• Cepstral-domain metrics are quantitative techniques that transform signals into cepstral features using methods like MFCC extraction and statistical aggregation.
• They apply signal processing steps such as windowing, FFT, DCT, and liftering to derive key measures like durable power components and temporal derivatives.
• These metrics support practical applications including speech authenticity, system identification, and anomaly detection through robust signal comparisons.

Cepstral-domain metrics are quantitative techniques that extract, summarize, and compare information from the cepstrum of a signal, typically used in speech processing, system identification, and dynamical time series analysis. The cepstrum—a nonlinear transformation involving the inverse Fourier transform of a logarithmic power (or magnitude) spectrum—facilitates the separation and measurement of periodic and resonant structures obscured in the frequency domain, enabling robust discrimination and system comparison tasks. Modern systems incorporate a variety of cepstral-domain statistics, including Mel-frequency cepstral coefficients (MFCCs), their temporal derivatives, statistical aggregations such as mean and variance, “durable power components” in low-quefrency regions, and weighted cepstral distances for measuring similarity between signals or underlying dynamical systems (Singh et al., 2020, Lauwers et al., 2018).

1. Mathematical Foundations of the Cepstrum and Cepstral Metrics

Given a time-domain signal $x[n]$, typically windowed and framed for short-time processing, the discrete Fourier transform (DFT) is computed: $$X[k] = \sum_{n=0}^{N-1} x[n]\,e^{-j\,2\pi kn/N}, \quad k=0,\dots,N-1.$$ The power cepstrum $c[q]$ is defined as the inverse DFT of the log power spectrum: $$c[q] = \frac{1}{N}\sum_{k=0}^{N-1} \log|X[k]|^2\,e^{j2\pi kq/N},\quad q=0,\dots,N-1.$$ The Mel-frequency cepstrum (MFCC) further applies a bank of overlapping triangular filters $H_m[k]$ on the Mel scale, log-compresses the filter energies, and computes a Type-II DCT: $$S_m = \sum_{k=0}^{N-1} |X[k]|^2 H_m[k], \quad \mathrm{MFCC}_n = \sum_{m=1}^M \log S_m \cos\left[\frac{\pi}{M} (n-\tfrac{1}{2})(m-\tfrac{1}{2})\right].$$ Framewise delta and delta-delta derivatives are constructed by finite temporal differencing: $$\Delta \mathrm{MFCC}_n[t] = \frac{\sum_{l=1}^L l\left(\mathrm{MFCC}_n[t+l]-\mathrm{MFCC}_n[t-l]\right)}{2\sum_{l=1}^L l^2},\quad \Delta^2 \mathrm{MFCC}_n[t] = \Delta\left(\Delta \mathrm{MFCC}_n[t]\right).$$ For deterministic single-input single-output (SISO) linear time-invariant (LTI) systems, the power cepstrum of the system transfer function $H(z)$, relating input $u[k]$ and output $y[k]$, is obtained by subtraction: $c_h(k) = c_y(k) - c_u(k),$ where $c_y(k)$ and $c_u(k)$ are the cepstra of output and input, respectively (Lauwers et al., 2018).

2. Signal Processing Steps and Metric Extraction

Standard signal-processing steps for cepstral metric extraction involve pre-emphasis, windowing (commonly 20–30 ms Hamming windows with 10 ms frame shift), DFT computation (FFT size $\approx 512$ or $1024$), and Mel-filterbank application (typically $M=20$–40 filters over 0–8 kHz). The DCT compresses information into the first 12–13 MFCCs per frame. Temporal derivatives extend the feature set with dynamic information.

Statistical aggregation is performed by computing, across all frames in an utterance: $$\mu_{\mathrm{MFCC}} = \frac{1}{T}\sum_{t=1}^T \overline{\mathrm{MFCC}}[t],\quad \sigma^2_{\mathrm{MFCC}} = \frac{1}{T} \sum_{t=1}^T (\overline{\mathrm{MFCC}}[t] - \mu_{\mathrm{MFCC}})^2,$$ where $\overline{\mathrm{MFCC}}[t]$ denotes the mean over MFCC channels, and similarly for delta and delta-delta statistics (Singh et al., 2020). Optional “liftering” can further bias the retained low-quefrency information: $$c_{\mathrm{lift}}[q] = \left(1+\frac{L}{2}\sin\frac{\pi q}{L}\right)c[q].$$

3. Specialized Cepstral Metrics: Durable Power Component and Weighted Cepstral Distance

The “durable power component” (DPC) is a scalar statistic representing the mean absolute cepstral energy in the lowest quefrency bins: $$DPC = \frac{1}{Q_0} \sum_{q=0}^{Q_0-1} |c[q]|, \quad Q_0 \text{ typically } 5\text{--}10,$$ used as an indicator of the vocal-tract structure in human speech, marked by a more prominent low-quefrency ridge relative to AI-synthesized speech (Singh et al., 2020).

The weighted cepstral distance between signals $y_1,u_1$ and $y_2,u_2$ is defined by

$$d_w(y_1,y_2) = \left[\sum_{k=0}^{\infty} w(k) (c_{h1}(k) - c_{h2}(k))^2\right]^{1/2},$$

with $w(k)\geq0$ (often $w(k)=k$ for $k\geq1$; $k=0$ excluded by convention). This metric admits closed-form expressions when model poles and zeros are known, and for ARMA models can be related to the Hilbert–Schmidt norm of an associated Hankel matrix. For stable, minimum-phase systems: $$\|H\|_c^2 =\log\frac{\prod_{i=1}^p\prod_{j=1}^q|1-\alpha_i\bar\beta_j|^2}{\bigl[\prod_{i,j=1}^p(1-\alpha_i\bar\alpha_j)\bigr]\;\bigl[\prod_{i,j=1}^q(1-\beta_i\bar\beta_j)\bigr]}.$$ This metric enables direct, data-driven comparison of linear dynamics, with geometric interpretation in the subspace angles between system observability matrices (Lauwers et al., 2018).

4. Empirical Distinctions and Interpretive Context

In digital audio forensics, Singh and Singh (Singh et al., 2020) demonstrate that the DPC and related MFCC metrics allow near-perfect discrimination between human and AI-synthesized speech. Key empirical findings include:

• Human speech exhibits a DPC roughly 30–50% higher than that of AI-synthesized voices; for $Q_0=5$, DPC averages $0.18$ (human) versus $0.10$ (AI), enabling $\sim90\%$ accuracy at a $0.14$ threshold.
• Mel spectrograms display intense low-quefrency ridges for humans, absent in AI speech.
• The first-order MFCC energy for human voices typically exceeds $10$ dB, while AI-synthesized voices fall below $7$ dB.
• Variance of $\Delta$ and $\Delta^2$ cepstral metrics is higher for human speech (approx. $0.12$) than for neural-synthesized outputs (approx. $0.07$).

Weighted cepstral distance, as extended by Geerts and Helsen (Lauwers et al., 2018), enables direct model-norm-based quantification of dynamical similarity, interpretable through the lens of pole-zero locations and subspace geometry. For pure minimum-phase or maximum-phase cases, subspace angle interpretations hold; for mixed-phase systems, the norm loses its direct geometric meaning but remains a computable metric.

5. Computational Workflow and Practical Implementation

Cepstral-domain metric extraction typically involves:

1. Preprocessing: Input conversion, segmentation, and windowing.
2. Spectral estimation: FFT-based PSD computation (Welch’s method is standard for time-series datasets).
3. Logarithmic compression and Mel scaling (as needed).
4. Cepstral computation: Inverse FFT of log-PSD, optional liftering.
5. Feature aggregation: Extraction of MFCCs, $\Delta$- and $\Delta^2$-coefficients, DPC, and statistical means/variances per utterance.
6. For distance metrics: Compute transfer-function cepstra via $c_h(k)=c_y(k)-c_u(k)$, apply prescribed weighting, and sum to obtain $d_w$.

Principal-angle-based geometric analysis requires construction of observability matrices and can be performed directly on projected Hankel matrices via SVD or eigenanalysis (Lauwers et al., 2018). Phase-type testing is accomplished by examining the sign support of the complex cepstrum after phase unwrapping, distinguishing minimum-phase (support on $k\geq 0$), maximum-phase (support on $k<0$), and mixed-phase cases.

6. Applications and Integration in Machine Learning Systems

Cepstral-domain metrics support a range of tasks:

• Speech authenticity detection: Aggregate cepstral features (mean/variance of MFCC, $\Delta$, $\Delta^2$, and DPC) combined with bispectral statistics enable SVM-based discrimination of human versus synthetic speech, achieving AUC $\approx 0.99$ and test accuracy up to $97.6\%$ in binary classification (Singh et al., 2020).
• Time-series clustering: Weighted cepstral distances group signals by underlying dynamical similarity without explicit system identification (Lauwers et al., 2018).
• Fault and anomaly detection: Online measurements can be rapidly compared to baseline cepstral descriptors to identify system changes.
• Texture and activity recognition: Cepstral metrics extract features from ARMA model representations of dynamic textures.
• Structural-health monitoring: Changes in cepstral distances signal shifting transfer function characteristics in engineered systems.

A plausible implication is that cepstral-domain approaches, owing to their compactness and interpretability, are well-suited for efficient, data-driven analytics in high-throughput environments where model-based parameterizations are impractical.

7. Limitations and Interpretive Boundaries

Cepstral metrics are constrained by the underlying assumptions of linearity, stationarity, and invertibility (for transfer-function approaches). In mixed-phase cases, the weighted cepstral distance loses strict dynamical interpretability, and certain pole-zero configurations become indistinguishable with respect to this metric (Lauwers et al., 2018). The durable power component's effectiveness as a discriminator is empirically robust for current neural speech synthesis models, but systematic improvement in synthesis quality may eventually reduce the empirical efficacy of such statistics, necessitating ongoing recalibration and multi-metric integration.

Cepstral-domain metrics, particularly in hybrid forms combining spectral, statistical, and bispectral information, remain fundamental to robust signal characterization, modeling, and comparison in speech science, audio forensics, and generalized dynamical system analysis.