What Is Cepstral Analysis?

Cepstral analysis is a nonlinear signal processing technique that applies two consecutive spectral transforms with a logarithmic operation between them, transforming a signal into a domain where convolutionally mixed components can be separated. The technique was introduced in 1963 by B. P. Bogert, M. J. Healy, and J. W. Tukey, who coined the term "cepstrum" by reversing the first four letters of "spectrum," and invented associated terminology, including "quefrency" (from "frequency") and "saphe" (from "phase"), to describe the properties of this new domain. The independent variable of the cepstrum is quefrency, measured in units of time, and its value corresponds to the period of a periodic structure present in the log spectrum.

The cepstrum is defined as the inverse Fourier transform of the logarithm of the power spectrum of a signal. In practice, the computation proceeds by windowing the time-domain signal, computing the discrete Fourier transform, taking the logarithm of the magnitude, and applying a second transform, yielding cepstral coefficients. The logarithm is the critical step: because convolution in the time domain corresponds to multiplication in the frequency domain, taking the log converts that multiplication into addition, allowing the two mixed components to be treated as separable additive contributions in the cepstral domain. The original formulation is detailed in the IEEE Signal Processing Magazine historical account by Oppenheim and Schafer.

Speech Analysis and Homomorphic Processing

The most consequential application of cepstral analysis in its early decades was in speech processing. Voiced speech is well modeled as the convolution of a glottal excitation signal (the periodic source produced by the vocal folds) with the vocal tract impulse response (the filter that shapes the spectral envelope). Separating these two components by conventional linear filtering is not feasible because they are convolutionally combined, not additively superimposed. Cepstral analysis provides a practical route to this separation through homomorphic signal processing, a broader theoretical framework developed by Alan Oppenheim at MIT in the 1960s that generalizes linear filtering to signals combined by nonlinear operations such as convolution. In the cepstral domain, the excitation contribution appears as a peak at a quefrency corresponding to the fundamental period (pitch), while the vocal tract contribution concentrates at low quefrencies, allowing the two to be separated by a simple windowing operation called liftering.

Mel-Frequency Cepstral Coefficients

For automatic speech recognition and music information retrieval, raw cepstral coefficients are rarely used directly. The mel-frequency cepstral coefficient (MFCC) representation, developed through the 1970s and 1980s, maps the frequency axis to the mel scale before computing the cepstrum, weighting lower frequencies more heavily in a way that approximates the frequency resolution of human auditory perception. As described in the Introduction to Speech Processing text from Aalto University, MFCCs are computed by applying a mel-filterbank to the power spectrum, taking the logarithm of each filter output, and then applying the discrete cosine transform. The result is a compact, decorrelated feature vector that captures the spectral shape of a speech frame efficiently. MFCCs became the dominant front-end feature representation for hidden Markov model (HMM) based speech recognition systems and remain widely used as input features for deep learning models in speech and audio tasks.

Applications

Cepstral analysis has applications across a range of signal processing and engineering domains, including:

• Automatic speech recognition as a feature extraction front-end
• Speaker identification and verification systems
• Pitch detection and fundamental frequency estimation in speech and music signals
• Music information retrieval, including instrument identification and genre classification
• Mechanical fault detection through cepstral analysis of vibration signals from rotating machinery
• Echo detection and removal in seismic and acoustic recordings