PLP coefficients describe speech spectral features compactly and speaker-independently

Evidence

Perceptual linear prediction (PLP) (Hermansky, 1990). Conceptually, the PLP coefficients and MFCCs are extracted in a similar manner. In PLP, however, the coefficients are extracted from a specific loudness spectrum. Moreover, the specific loudness spectra are transformed back into autocorrelation functions using the inverse Fourier transform, after which the traditional autocorrelation-based LP algorithm is used to extract the coefficients. The resulting coefficients are able to describe the spectral features of speech signals in a compact and rather speaker-independent manner. — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 1599

Categorization

• SEED: The trick of bouncing from loudness spectrum back to autocorrelation via inverse FFT before running LP is a genuinely elegant technical maneuver — a piece for audio-tech readers on how perception-aware front-ends work.
• QUESTION: Why does transforming loudness spectra back into autocorrelation functions improve speaker-independence? Worth chasing the underlying mechanism.
• XREF: connects to: MFCCs remain the dominant feature representation for speech recognition systems — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 1593