spektralcentroid

Spektralcentroid, more commonly known as the spectral centroid, is a measure used in digital signal processing and audio analysis to characterize a signal’s frequency content. It represents the center of gravity of the magnitude spectrum of a windowed signal and is often interpreted as the perceived brightness or timbre of the sound.

Calculation is typically performed on a short-time Fourier transform magnitude spectrum. For a windowed frame, let M[k] be the magnitude at frequency f[k] = k * Fs / N, where Fs is the sampling rate and N is the FFT size. The spektralcentroid is given by C = sum_k f[k] * M[k] / sum_k M[k]. Some variants use the power spectrum, or different weighting schemes, but the general idea is a weighted average of the spectrum’s frequencies.

Interpretation and use are straightforward: higher centroid values indicate more energy at higher frequencies, associated with brighter, crisper timbres; lower centroid values indicate energy concentrated at lower frequencies, associated with darker sounds. The feature is widely used in music information retrieval and audio analysis to aid tasks such as instrument recognition, genre classification, tempo-independent timbre analysis, and audio thumbnailing. It is often combined with other spectral descriptors like bandwidth, rolloff, and MFCCs to build more robust models.

Limitations include sensitivity to spectral resolution, window length, and noise, which can shift centroid estimates without reflecting true perceptual changes. It also cannot capture multiple concurrent tonal components or harmonic structure on its own, so it is typically used in conjunction with additional features.