melband

Melband is a term most commonly encountered in the field of audio signal processing, where it refers to a frequency band defined on the mel scale. The mel scale is a perceptual frequency scale in which equal steps correspond to roughly equal perceived pitch differences. In practice, a mel-filter bank consists of overlapping triangular filters that partition the spectrum into a set of mel bands.

To obtain mel-band energies, a typical pipeline is used. The audio signal is divided into short frames, and a Fourier transform yields the magnitude spectrum. The magnitude (or power) spectrum is then passed through the mel-filter bank, producing one energy value per mel band. Often the logarithm of these energies is taken, and in many systems the log-mel energies are further transformed with a discrete cosine transform to compute Mel-Frequency Cepstral Coefficients (MFCCs).

The number of mel bands and the frequency range are common design choices. Typical configurations use anywhere from about 20 to 128 mel bands, with a frequency range tailored to the application (for example, 0–8 kHz for telephony, broader ranges for music and general audio analysis).

Applications of mel-band features are widespread. They are central to automatic speech recognition, speaker identification, and music information retrieval, and they are also used in acoustic scene analysis and various noise-robust audio processing tasks. Variants include direct mel-band energy features and their use as input to classification or regression models, as well as MFCCs, which add a decorrelation step via the DCT.

In summary, a melband is a perceptually spaced spectral band used to describe or extract energies from audio signals for a variety of analyses and recognition tasks. See also mel scale, mel-filter bank, MFCCs.