Logmel

Logmel, short for log-mel spectrogram, is a time-frequency representation used in audio processing and machine learning. It is derived from an audio waveform by computing a spectrogram with a mel frequency scale and applying logarithmic compression to the magnitude (or power) values. The result highlights perceptually relevant energy patterns across time and frequency while reducing dynamic range.

Computation typically proceeds as follows: compute the short-time Fourier transform to obtain a magnitude spectrum; apply a mel filter bank to warp frequencies onto the mel scale, producing a mel-spectrogram; apply log compression, often using log(x + epsilon), to obtain the log-mel spectrogram. The final output is a two-dimensional matrix with time frames along one axis and mel bands along the other.

Common parameter choices include a sampling rate of 16 kHz or 22.05 kHz, a window length of about 20–25 milliseconds, a hop length around 10 milliseconds, and a number of mel bands typically in the range of 40–128. The mel filter bank is often defined over a frequency range from 0 up to either a chosen max frequency or the Nyquist limit, depending on the application. The exact design influences perceptual fidelity and downstream performance.

Logmel features are widely used as inputs to neural networks for tasks such as automatic speech recognition, speaker identification, and music tagging. They are related to MFCCs, which are derived by applying a discrete cosine transform to the log-mel energies; thus logmel preserves more local spectral information suitable for end-to-end modeling. Common implementations are available in libraries such as Librosa and torchaudio.