harmonicpercussive

Harmonicpercussive, in the context of audio signal processing and music information retrieval, refers to the separation of an audio signal into two complementary components: harmonic content and percussive content. The harmonic component consists of sustained, pitched elements such as notes from melodies or harmonies, while the percussive component comprises transient, broadband events like drums or plucked string attacks. When used as a formal technique, it is often described as harmonic-percussive source separation (HPSS).

Most HPSS methods operate on the magnitude of the short-time Fourier transform (STFT) of an audio signal. The standard approach estimates two spectrogram components: a harmonic estimate that emphasizes energy persistent over time, and a percussive estimate that emphasizes energy spread across frequencies in short time frames. The two estimates are combined with soft masks to separate the signal into harmonic and percussive spectrograms, which are then converted back to time-domain signals using the original phase information. Variants may apply median filtering along time or frequency axes to isolate the respective components.

Applications of harmonicpercussive separation include music remixing, source separation for improved transcription or alignment, denoising, and feature extraction in music information retrieval. It is commonly implemented in audio analysis libraries and software, with both traditional DSP-based methods and modern deep learning approaches that predict masks or directly estimate component spectra. Limitations arise when harmonic and percussive elements overlap in time and frequency, leading to leakage or artifacts. HPSS remains a useful preprocessing step but is seldom perfect for complex or highly dense mix content. See also source separation and music information retrieval.