Cepstral Transform Overview

• Cepstral transform is a homomorphic spectral representation that converts convolution in the time domain into additive structures in the quefrency domain.
• It underpins practical applications such as MFCC-based speech recognition, seismic deconvolution, and random-field spectral analysis.
• Recent developments include fractional, complex, and multi-layered variants that enhance computational efficiency and signal feature extraction.

The cepstral transform is a homomorphic representation obtained by applying a logarithm to a spectral representation and then transforming back to a delay-like domain usually called the quefrency domain. In its classical real form for a discrete signal $x(n)$, it is written as $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$, while for a stationary process it is the inverse Fourier transform of the log power spectrum. This construction converts spectral multiplication into addition and thereby supports deconvolution, source–filter separation, spectral parametrization, and statistically motivated smoothing in applications ranging from speech and seismic analysis to random fields and Green–Kubo transport estimation (Miah et al., 2012).

1. Classical formulation and homomorphic principle

For a discrete-time signal $x(n)$ with spectrum $X(\omega)=\mathcal{F}\{x(n)\}$, the real cepstrum may be written in liftered form as

$c(n)=\mathcal{F}^{-1}\{H(\omega)\log|X(\omega)|\},$

where $H(\omega)$ is a frequency-domain mask and the independent variable $n$ is interpreted as quefrency rather than time. In the corresponding stochastic-process formulation, if $\Phi(e^{j\omega})>0$ denotes a power spectral density, then

$$c[n]=\frac{1}{2\pi}\int_0^{2\pi} e^{jn\omega}\log \Phi(e^{j\omega})\,d\omega, \qquad \log \Phi(e^{j\omega})=\sum_{n=-\infty}^{\infty} c[n]e^{-jn\omega}.$$

These expressions make explicit that the cepstrum is the Fourier-series representation of the log spectrum (Yu et al., 2019).

The central homomorphic property is that convolution in the time domain becomes addition after logarithmic spectral mapping. In speech and music, $|X(\omega)|\approx |E(\omega)|\,|V(\omega)|$ yields

$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$0

and the inverse transform places the excitation and envelope contributions in different quefrency regions. In seismic analysis, the same principle is written as $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$1, implying $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$2, hence $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$3, and therefore $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$4. This additive separation underlies homomorphic deconvolution and wavelet estimation (Miah et al., 2012).

The quefrency interpretation is operational rather than merely terminological. High-quefrency peaks correspond to periodicities such as a fundamental period, whereas low-quefrency components capture the slowly varying spectral envelope. Liftering is filtering in this quefrency domain: low-quefrency liftering extracts envelope structure, while high-quefrency liftering isolates periodic excitation or pitch-related content (Yu et al., 2019).

2. MFCC as a cepstral implementation in speech analysis

A widely used speech-processing realization of the cepstral idea is the Mel Frequency Cepstral Coefficient pipeline. In the formulation described for voice recognition, the signal first undergoes pre-emphasis,

$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$5

followed by framing over short quasi-stationary intervals, windowing with a Hamming window,

$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$6

and a DFT,

$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$7

The periodogram $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$8 is then passed through a triangular mel filterbank whose center frequencies are defined using

$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$9

Mel-band energies are formed as

$x(n)$0

followed by logarithmic compression $x(n)$1 and a type-II DCT,

$x(n)$2

Typical implementation values in the walkthrough include $x(n)$3, $x(n)$4, and retention of $x(n)$5 coefficients often equal to $x(n)$6–$x(n)$7 (Muda et al., 2010).

In this pipeline, the logarithm compresses the dynamic range of spectral energies, the mel warp gives greater emphasis to low frequencies, and the DCT of log-energies approximates the cepstrum. The stated effect is to separate fast-varying pitch structure from the slowly varying spectral envelope and to retain chiefly the envelope of the log-mel spectrum, which carries most of the phonetic information while discarding fine pitch harmonics and noise. The cited work combines MFCC extraction with Dynamic Time Warping as the feature-matching stage, using MFCC as a non-parametric model of the human auditory perception system and DTW for nonlinear temporal alignment in voice recognition (Muda et al., 2010).

3. Complex cepstrum and mixed-phase decomposition

The complex cepstrum extends the real cepstrum by retaining phase information through the complex logarithm. For a windowed speech segment $x(n)$8 with DTFT $x(n)$9, it is defined as

$X(\omega)=\mathcal{F}\{x(n)\}$0

with

$X(\omega)=\mathcal{F}\{x(n)\}$1

where $X(\omega)=\mathcal{F}\{x(n)\}$2 is the unwrapped phase. In practice, the computation uses a fine-grid FFT, magnitude and phase extraction, one-dimensional phase unwrapping, formation of $X(\omega)=\mathcal{F}\{x(n)\}$3, and an inverse FFT to the quefrency domain (Drugman et al., 2019).

Its main structural use is causal–anticausal decomposition. The complex cepstrum is split into

$X(\omega)=\mathcal{F}\{x(n)\}$4

and exponentiation of the two separated cepstra reconstructs two spectral factors $X(\omega)=\mathcal{F}\{x(n)\}$5 and $X(\omega)=\mathcal{F}\{x(n)\}$6 satisfying $X(\omega)=\mathcal{F}\{x(n)\}$7. The nonnegative-quefrency part corresponds to a minimum-phase causal factor; the negative-quefrency part corresponds to a maximum-phase anticausal factor. In mixed-phase speech modeling, the anticausal component is associated with the glottal open phase, whereas the causal part combines the glottal return phase and the vocal-tract response (Drugman et al., 2019).

For glottal source estimation, the cited studies make the windowing conditions explicit. Frames must be centered on a Glottal Closure Instant, with error larger than $X(\omega)=\mathcal{F}\{x(n)\}$8 ms degrading performance. A two-period generalized cosine window

$X(\omega)=\mathcal{F}\{x(n)\}$9

is recommended, with optimal length approximately $c(n)=\mathcal{F}^{-1}\{H(\omega)\log|X(\omega)|\},$0 for Hanning and approximately $c(n)=\mathcal{F}^{-1}\{H(\omega)\log|X(\omega)|\},$1 for Blackman. The papers state that if these rules are violated, for example by using a $c(n)=\mathcal{F}^{-1}\{H(\omega)\log|X(\omega)|\},$2 ms Hanning window, zeros inside and outside the unit circle do not separate and the decomposition fails (Drugman et al., 2019).

A principal motivation for the complex cepstral route is computational. One comparison states that ZZT decomposition requires polynomial root-finding with cost roughly $c(n)=\mathcal{F}^{-1}\{H(\omega)\log|X(\omega)|\},$3 for an $c(n)=\mathcal{F}^{-1}\{H(\omega)\log|X(\omega)|\},$4-tap frame, while complex cepstrum requires FFT/IFFT operations and phase unwrapping with cost $c(n)=\mathcal{F}^{-1}\{H(\omega)\log|X(\omega)|\},$5. In another formulation, ZZT is described as $c(n)=\mathcal{F}^{-1}\{H(\omega)\log|X(\omega)|\},$6 in practice, whereas the complex cepstrum remains $c(n)=\mathcal{F}^{-1}\{H(\omega)\log|X(\omega)|\},$7. Reported Matlab timings at $c(n)=\mathcal{F}^{-1}\{H(\omega)\log|X(\omega)|\},$8 kHz include approximately $c(n)=\mathcal{F}^{-1}\{H(\omega)\log|X(\omega)|\},$9 ms/frame for ZZT versus approximately $H(\omega)$0 ms/frame for the complex-cepstrum approach at $H(\omega)$1 Hz, and relative timings of $H(\omega)$2 versus $H(\omega)$3 at $H(\omega)$4 Hz and $H(\omega)$5 versus $H(\omega)$6 at $H(\omega)$7 Hz (Drugman et al., 2019, Drugman et al., 2019).

4. Generalized and recursive variants

Several later formulations preserve the log-spectrum-to-quefrency logic while altering the underlying transform, the recursion, or the spectral domain on which the cepstrum is defined.

Variant	Defining transform	Reported use
Fractional-cepstrum	$H(\omega)$8	Seismic wavelet estimation; multidomain analysis
Multi-layered cepstrum	Recursive Fourier, high-pass mask, and nonlinear activation	MF0 estimation under contamination
Pseudo-cepstrum	DCT of pseudo-inverse-mel log spectrum	Pitch modification for mel-based vocoders

The fractional-cepstrum replaces the Fourier transform with the fractional Fourier transform. For order $H(\omega)$9,

$n$0

Because $n$1 interpolates continuously between the identity operator at $n$2 and the ordinary Fourier transform at $n$3, the resulting representation provides a family of cepstral views. The stated advantages are enhanced separation of overlapping components for chirp-like or time–frequency-varying signals, tunable resolution by choosing $n$4, and multidomain analysis over several fractional orders. In the synthetic seismic example, real fractional-cepstra were computed for $n$5, and the prominence and separation of quefrency peaks changed with $n$6. The paper notes that extension to the complex fractional-cepstrum will involve phase unwrapping and the FRFT’s extra chirp-phase term (Miah et al., 2012).

The multi-layered cepstrum iterates the cepstral mechanism recursively. With $n$7 as the DFT matrix and $n$8 as a diagonal high-pass mask, it is defined by

$n$9

where $\Phi(e^{j\omega})>0$0 is a power-scaled rectifier. Even layers apply masking in frequency, odd layers in quefrency. The intended effect is progressive suppression of slowly varying or aperiodic components and progressive enhancement of periodic structure. On the Bach10 dataset, reported F-scores increase from approximately $\Phi(e^{j\omega})>0$1 for the $\Phi(e^{j\omega})>0$2-layer generalized cepstrum to approximately $\Phi(e^{j\omega})>0$3 for the $\Phi(e^{j\omega})>0$4-layer configuration; under severe high-pass degradation, the $\Phi(e^{j\omega})>0$5-layer system retains approximately $\Phi(e^{j\omega})>0$6 versus approximately $\Phi(e^{j\omega})>0$7 for the $\Phi(e^{j\omega})>0$8-layer case (Yu et al., 2019).

The pseudo-cepstrum is defined not on a classical linear-frequency spectrum but on a mel representation. The method first constructs $\Phi(e^{j\omega})>0$9 through a pseudo-inverse mel transform $$c[n]=\frac{1}{2\pi}\int_0^{2\pi} e^{jn\omega}\log \Phi(e^{j\omega})\,d\omega, \qquad \log \Phi(e^{j\omega})=\sum_{n=-\infty}^{\infty} c[n]e^{-jn\omega}.$$0, then applies a DCT: $$c[n]=\frac{1}{2\pi}\int_0^{2\pi} e^{jn\omega}\log \Phi(e^{j\omega})\,d\omega, \qquad \log \Phi(e^{j\omega})=\sum_{n=-\infty}^{\infty} c[n]e^{-jn\omega}.$$1 The cepstral peak is shifted in quefrency according to a desired semitone shift, and the modified mel representation is reconstructed by inverse DCT and mel filtering. The paper emphasizes that this is not a classical cepstrum of the time-domain signal, because the pseudo-inverse mel transform is inserted before the DCT. The reported advantage is compatibility with any mel-based vocoder without additional training or architectural change (Ellinas et al., 18 Dec 2025).

5. Multidimensional, generalized-moment, and random-field cepstra

For multidimensional stationary fields, the cepstrum becomes a Fourier-series expansion of a multidimensional log spectrum. If $$c[n]=\frac{1}{2\pi}\int_0^{2\pi} e^{jn\omega}\log \Phi(e^{j\omega})\,d\omega, \qquad \log \Phi(e^{j\omega})=\sum_{n=-\infty}^{\infty} c[n]e^{-jn\omega}.$$2 is the power spectrum of a $$c[n]=\frac{1}{2\pi}\int_0^{2\pi} e^{jn\omega}\log \Phi(e^{j\omega})\,d\omega, \qquad \log \Phi(e^{j\omega})=\sum_{n=-\infty}^{\infty} c[n]e^{-jn\omega}.$$3-dimensional random field, then the classical one-dimensional relation generalizes to multidimensional cepstral moments. In one formulation, for $$c[n]=\frac{1}{2\pi}\int_0^{2\pi} e^{jn\omega}\log \Phi(e^{j\omega})\,d\omega, \qquad \log \Phi(e^{j\omega})=\sum_{n=-\infty}^{\infty} c[n]e^{-jn\omega}.$$4 the generalized moments reduce to

$$c[n]=\frac{1}{2\pi}\int_0^{2\pi} e^{jn\omega}\log \Phi(e^{j\omega})\,d\omega, \qquad \log \Phi(e^{j\omega})=\sum_{n=-\infty}^{\infty} c[n]e^{-jn\omega}.$$5

whereas for $$c[n]=\frac{1}{2\pi}\int_0^{2\pi} e^{jn\omega}\log \Phi(e^{j\omega})\,d\omega, \qquad \log \Phi(e^{j\omega})=\sum_{n=-\infty}^{\infty} c[n]e^{-jn\omega}.$$6 the moments involve $$c[n]=\frac{1}{2\pi}\int_0^{2\pi} e^{jn\omega}\log \Phi(e^{j\omega})\,d\omega, \qquad \log \Phi(e^{j\omega})=\sum_{n=-\infty}^{\infty} c[n]e^{-jn\omega}.$$7. Given finite covariance data $$c[n]=\frac{1}{2\pi}\int_0^{2\pi} e^{jn\omega}\log \Phi(e^{j\omega})\,d\omega, \qquad \log \Phi(e^{j\omega})=\sum_{n=-\infty}^{\infty} c[n]e^{-jn\omega}.$$8 and generalized cepstral data $$c[n]=\frac{1}{2\pi}\int_0^{2\pi} e^{jn\omega}\log \Phi(e^{j\omega})\,d\omega, \qquad \log \Phi(e^{j\omega})=\sum_{n=-\infty}^{\infty} c[n]e^{-jn\omega}.$$9, the paper formulates a generalized maximum-entropy problem. Its main theorem states that if the covariance data admit some positive spectrum on an open set and $|X(\omega)|\approx |E(\omega)|\,|V(\omega)|$0, then for any $|X(\omega)|\approx |E(\omega)|\,|V(\omega)|$1 there is a unique interior solution to the regularized dual problem, producing a rational spectrum

$|X(\omega)|\approx |E(\omega)|\,|V(\omega)|$2

that matches covariances exactly and the generalized cepstral coefficients approximately, with error controlled by $|X(\omega)|\approx |E(\omega)|\,|V(\omega)|$3. When spectral factorization is feasible, a shaping filter can then be constructed from $|X(\omega)|\approx |E(\omega)|\,|V(\omega)|$4 and $|X(\omega)|\approx |E(\omega)|\,|V(\omega)|$5 (Zhu et al., 2021).

In two-dimensional random-field modeling, the spectral density $|X(\omega)|\approx |E(\omega)|\,|V(\omega)|$6 is represented as

$|X(\omega)|\approx |E(\omega)|\,|V(\omega)|$7

with cepstral coefficients

$|X(\omega)|\approx |E(\omega)|\,|V(\omega)|$8

The cited theory provides recursive formulas linking the cepstral coefficients to an equivalent moving-average random field, facilitating computation of the autocovariance matrix. A central structural property is that because $|X(\omega)|\approx |E(\omega)|\,|V(\omega)|$9 is strictly positive on the torus, the induced covariance is positive definite without stability or invertibility constraints on the cepstral coefficients. Under expanding-domain asymptotics, Bayesian, maximum likelihood, and quasi-maximum likelihood estimators are analyzed, and in the special cepstral model the Fisher information for coefficients other than $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$00 is the identity matrix, with asymptotic variance $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$01 and vanishing cross-covariances among distinct coefficients (McElroy et al., 2011).

These formulations place the cepstrum within spectral moment problems rather than only within signal decomposition. In that setting, cepstral coefficients act as unconstrained or generalized log-spectral coordinates that can be combined with covariance constraints to produce rational spectral estimators and shaping filters (Zhu et al., 2021).

6. Statistical estimation, robust inference, and non-speech applications

In stationary time-series classification, the cepstrum is used as a low-dimensional representation of the log spectrum. One robust construction begins with a multitaper periodogram and an M-periodogram, defines

$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$02

and then applies Linear Discriminant Analysis to the resulting coefficient vectors. The stated advantages are that the zeroth coefficient isolates white-noise variance, higher-order coefficients capture autocorrelation structure, and robust spectral estimation improves resistance to outliers and heavy-tailed innovations. The exposition notes that the first few coefficients often provide a parsimonious representation, with examples such as $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$03 or $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$04 (Matias et al., 2024).

A related semiparametric model for replicated covariate-dependent time series represents each replicate-specific log spectrum $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$05 by a truncated cepstral expansion,

$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$06

then estimates the coefficients in two stages. Stage 1 uses a Whittle likelihood-based minimum-contrast estimator for the truncated cepstral vector $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$07; Stage 2 fits a multivariate linear model

$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$08

to relate cepstral coefficients to covariates. The paper states that choosing $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$09 small, for example $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$10–$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$11, yields substantial computational savings. In the Parkinson’s-disease postural-control example, the data consist of $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$12 subjects and $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$13 observations per series, and bootstrap pointwise confidence intervals are used for frequency-varying covariate effects (Li et al., 2024).

A recent materials application uses cepstral analysis to stabilize Green–Kubo thermal-conductivity calculations for metal-organic frameworks. Starting from one Cartesian component $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$14 of the microscopic heat flux, the method computes the HFACF $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$15, the power spectrum

$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$16

the log spectrum $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$17, and discrete cepstral coefficients by inverse Fourier or cosine transform. Low-order coefficients encode slowly varying physical spectral features, while high-order coefficients capture rapidly oscillating noise-like fluctuations; truncation at order $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$18 is selected by minimizing the corrected Akaike criterion. For MOF-5, HKUST-1, and ZIF-8, using machine-learned moment tensor potentials trained on DFT reference data, the reported result is stable convergence within about $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$19–$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$20 ns of total sampling time, whereas direct Green–Kubo analysis required $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$21–$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$22 ns for comparable precision. The paper reports agreement with experiment to within $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$23, including the example $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$24 W m$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$25 K$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$26 versus $c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$27 W m$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$28 K$c(n)=\mathcal{F}^{-1}\{\ln|\mathcal{F}\{x(n)\}|\}$29 for MOF-5 (Lindner et al., 11 Jun 2026).

Across these uses, the cepstral transform serves as a coordinate system for log-spectral structure. In speech it supports source–filter decomposition and perceptually motivated feature extraction; in multidimensional spectral theory it provides generalized moment constraints and unconstrained spectral parametrizations; and in modern statistical workflows it functions as a compact, often robust representation for inference, regression, classification, and noise suppression (Matias et al., 2024).