Pre-emphasis
A preprocessing step, commonly implemented as a first-order high-pass difference filter, that relatively boosts higher-frequency components of speech.
Classical speech pipelines use pre-emphasis to partially flatten spectral tilt and give higher-frequency components more weight during feature extraction. A common form is y[n] = x[n] - αx[n-1], but the coefficient is not a universal constant.
Modern learned models should not be given this stage automatically; inference preprocessing must match the model's training pipeline. It is therefore a traditional DSP choice in MFCC feature extraction, while for Whisper compatibility with the expected frontend is the controlling rule.