Log‑mel Spectrogram

A log‑mel spectrogram is a way of turning a sound waveform into an image‑like picture that captures how energy is spread over time and pitch. First the signal is broken into tiny overlapping slices; for each slice a short‑time Fourier transform measures the strength of many evenly spaced frequencies. Those frequency bins are then regrouped onto the mel scale, a spacing that mirrors how the human ear perceives pitch, so low frequencies stay detailed while high frequencies are compressed. Finally the amplitudes are transformed with a logarithmic curve, which squashes very loud parts and lifts quiet ones, making the visual pattern more uniform and easier for algorithms to work with.

Because it mirrors both the frequency resolution of our hearing and the way we perceive loudness, the log‑mel spectrogram has become the standard front end for many audio‑based machine‑learning tasks. Speech recognizers use it to turn spoken words into patterns a neural network can decode; music genre classifiers read it to separate styles; voice conversion and synthesis systems employ it as a reversible representation that can be edited or regenerated. In any application where computers need to understand, generate, or compare sounds, the log‑mel spectrogram provides a compact yet perceptually meaningful snapshot of the audio.