Melscaled

Melscaled refers to data that has been transformed to the mel scale, a perceptual frequency scale used in audio processing. The mel scale is designed so that equal distances correspond to roughly equal perceived pitch differences, reflecting how humans hear frequencies. A common formulation maps frequency f in hertz to mel units with f_mel ≈ 2595 log10(1 + f/700).

In practice, a melscaled representation is produced by applying a bank of triangular mel-spaced filters to the magnitude or power spectrum obtained from a short-time Fourier transform. The result is a mel-spectrogram, where each coefficient corresponds to energy within a particular mel band. Taking the logarithm of these energies yields a log-mel spectrogram, which aligns better with human loudness perception and is widely used as input to neural networks.

Mel-scaled features are widely used in speech recognition, speaker identification, and music information retrieval. They reduce dimensionality and emphasize lower frequencies, which often contain important information for audio tasks. The approach is also favored for its balance between perceptual relevance and computational efficiency.

Typical practice uses a range of mel bands, commonly from about 40 to 128, with frame lengths around 20 to 40 milliseconds and hop lengths near 10 milliseconds. The mel scale can be applied at various sampling rates (for example 16 kHz or 44.1 kHz), and mel energies are often used to compute higher-level features such as Mel-frequency cepstral coefficients (MFCCs) by applying a discrete cosine transform.

See also: Mel scale, mel filter bank, mel-spectrogram, MFCCs, STFT.