Log-Mel Spectrogram

Turkish equivalent: Log-Mel spektrogramDomain: Speech Processing

A time-frequency representation formed by applying a Mel-scaled filter bank to a spectrum and taking logarithmic energies, widely used as speech-model input.

A log-Mel spectrogram is a time–frequency representation obtained by accumulating short-time spectral power in perceptually spaced Mel bands and applying logarithmic compression to each band energy.

Mathematical construction

A sampled signal x[n] is divided into overlapping frames using a window w[n] and hop H. If the Fourier transform for frame t is Xₜ[k], its power spectrum can be expressed as |Xₜ[k]|². For Mel-filter weights Hₘ[k], band energy is Eₘ(t) = Σₖ Hₘ[k]|Xₜ[k]|², followed by Lₘ(t) = log(max(Eₘ(t), ε)). The positive floor ε specifies numerical behavior for very small band energies.

Frame and hop length establish time resolution, FFT size defines the spectral sampling grid, and the number of Mel bands sets the frequency dimension of the resulting feature tensor. Filter normalization, frequency endpoints and magnitude-versus-power input are also part of the feature contract.

Relationship to MFCC

MFCC applies a discrete cosine transform to log-Mel band energies and retains a selected set of cepstral coefficients. Log-Mel features preserve the filter-bank axis; MFCC expresses the result in a cepstral basis. See Mel Scale, Filter Bank and Short-Time Fourier Transform.

Related technical publications

Publications whose title or summary directly references this concept.

Whisper Architecture in Speech Recognition Systems

Whisper converts 16 kHz audio to a log-Mel spectrogram and generates tokens autoregressively with an encoder-decoder Transformer; language, task, and timestamp information share the same token sequence.