librosa

Nearly every open source audio machine learning project depends on librosa somewhere in its pipeline. The package provides reference Python implementations for audio loading and resampling, short-time Fourier transforms, mel spectrograms, constant-Q transforms, beat and tempo tracking, onset detection, pitch estimation, harmonic-percussive separation, and the standard features like MFCCs and chroma used to feed models. Built on NumPy and SciPy, the API favors plain arrays over custom objects, which is why it composes so easily with PyTorch and scikit-learn workflows, and everything runs on CPU with no GPU or network access needed after a pip or conda install. Development continues as an active community project with around 8,500 GitHub stars, versioned documentation at librosa.org, and an accompanying SciPy conference paper as the standard citation. The code is ISC licensed, a permissive license equivalent in effect to MIT, so commercial use is unrestricted. For feature extraction, dataset preparation, and music information retrieval research it remains the default tool more than a decade after its first release.