AFX¶

AFX turns audio clips into fixed-length feature vectors for machine learning. It implements 47 features (time-domain, autocorrelation, spectral, spectrogram, cepstral, harmonic, linear-prediction, and the eGeMAPS v02 parameter set) with NumPy and SciPy, selects and parameterizes them from a JSON config, and aggregates the per-frame values into one vector per clip. The same config drives a scikit-learn transformer, a parallel batch extractor with a CLI, and (for 20 features) a differentiable PyTorch module.

Install¶

pip install afx-audio            # numpy, scipy, soundfile
pip install "afx-audio[all]"     # + scikit-learn, torch, pandas/pyarrow, matplotlib

The import name is afx_audio (import afx_audio as afx). Extras: sklearn, torch, parquet, plot, all, and dev for tests and docs.

One clip¶

import afx_audio as afx

config = afx.load_preset('speech@1')
signal, sr = afx.load_audio('clip.wav', sr=config['sample_rate'])
features = afx.extract_all_features(signal, sr, config)
names, vector = afx.flatten_features(features)

features maps names such as mfcc_mean to arrays; flatten_features turns it into column names and one float vector. For a long recording, names, X, starts = afx.extract_segments(signal, sr, config, segment_length=1.0, hop=0.5) gives one such vector per segment as the rows of X, with each segment's start time in seconds.

Single features are plain functions:

mfcc = afx.FEATURE_MAP['mfcc'](signal, sr, n_mfcc=20, frame_size=400, hop_length=160)['mfcc']
afx.feature_info('mfcc')        # group, units, parameters and defaults, definition
afx.list_features('spectral')

A folder¶

afx dataset/ --preset speech --output features.parquet --jobs 8

writes one row per file, keyed by the path relative to dataset/. Files that fail to load are listed in features.parquet.errors.json, and the Parquet metadata records the config hash, preset id, library versions and input file hashes (Reproducibility).

Where to go next¶

Configuration covers framing, aggregation, coverage, presets and the config schema; Features lists every feature with its units, parameters and undefined cases; eGeMAPS compares the parameter set with openSMILE; librosa equivalence lists which features reproduce librosa exactly and where AFX differs; Notebooks is the tutorial.