A quantitative metric that measures how uniformly power is distributed across a frequency band, commonly defined as the ratio of the geometric mean to the arithmetic mean of the power spectral density values over that band (often expressed in linear units or dB). Values near one (or 0 dB) indicate noise‑like (flat) spectra; values near zero indicate tone‑like (peaky) spectra.

Principle

Because the geometric mean is sensitive to multiplicative imbalance while the arithmetic mean captures total power, their ratio discriminates between spectrally flat (broadband/noise‑like) and spectrally peaky (tone‑dominated) signals independent of absolute power scaling when computed over a specified band and resolution.

Demonstration

Situation: Compute spectral flatness for a 1‑s audio sample using a PSD estimate over 512 frequency bins. Recognition: Pure tone produces a low flatness (close to 0), white noise produces flatness near 1. Action: Use flatness thresholding to decide compression or synthesis strategy. Consequence: Threshold‑based decisions separate tonal content from noise‑like content for coding or texture analysis, subject to windowing and spectral resolution choices.

Misapplication

Applying spectral flatness computed over an inappropriate bandwidth, insufficient frequency resolution, or on strongly nonstationary signals and treating the numeric value as a perceptual measure; or failing to account for window leakage and zero‑padding effects that bias the geometric mean.

Consequence

Properly used, the measure aids audio coding, source separation and feature extraction by distinguishing tonal from noise components; misused, it can misclassify signals and lead to suboptimal compression or incorrect perceptual inferences.

Reversal

For signals with both strong tones and broad noise floor, or for harmonically structured signals with many resolved partials, spectral flatness can yield intermediate values that are ambiguous; alternative or complementary metrics (spectral entropy, tonal/noise decomposition) may be required.

Boundary

Clearly within: stationary or locally stationary signals where PSD estimation over a chosen band and resolution is meaningful. Boundary case: short, transient signals where spectral estimates have high variance. Clearly outside: single‑snapshot narrowband measurements with inadequate bins or fundamentally nonstationary data without time‑frequency treatment.

Semantic Tension

Tension between using a single scalar flatness measure (global uniformity) and richer descriptors (spectral entropy, spectral centroid, time‑frequency features) that capture different aspects of signal structure; choice depends on task specificity.

Synthesis

Spectral flatness is a compact, scale‑invariant indicator of uniformity versus peaking in a spectrum; it is valuable for automated tonal/noise discrimination but must be computed with appropriate PSD estimation parameters and interpreted alongside complementary features.