Melspektrogram

A mel spectrogram, sometimes written as mel-spectrum or melspektrogram, is a time–frequency representation of an audio signal in which the frequency axis is mapped onto the mel scale. The mel scale is a perceptual scale that approximates how humans perceive pitch, placing more resolution at lower frequencies where auditory sensitivity is higher. This makes the mel spectrogram a compact and perceptually relevant alternative to a conventional spectrogram.

To compute a mel spectrogram, the audio signal is divided into short overlapping frames and transformed with the short-time Fourier transform to obtain magnitude spectra. The magnitude (or power) spectra are then passed through a bank of triangular mel filters, producing a set of mel-band energies for each frame. The resulting matrix can be used directly as a mel spectrogram or transformed by taking the logarithm of the energies, yielding a log-mel spectrogram. The log scale helps compress dynamic range and aligns with perceived loudness.

The mel axis is typically composed of a fixed number of bands, commonly ranging from about 40 to 128, spaced nonlinearly according to the mel scale. The time axis reflects the frame rate determined by the window length and hop size. Unlike complex spectrograms, mel spectrograms usually discard phase information, focusing on amplitude content for further analysis.

Applications include speech recognition, speaker identification, music information retrieval, and general audio classification. They serve as a robust, perceptually grounded feature set for machine learning models. Variants include log-mel spectrograms and power mel spectra. Limitations include dependence on STFT parameters and the loss of phase information, which can affect synthesis or precise pitch reconstruction. See also spectrogram, MFCC, and short-time Fourier transform.