AREA-Beamforming

by Marc C. Green

A clustering approach to sound source tracking in Ambisonic audio. This modules contains code for:

• Spherical harmonic eamforming using Plane Wave Decomposition and Cross-pattern coherence beams.
• Rotation of non axis-symmetric spherical functions using Wigner-D matrices.
• Fibonacci, regular and geodesic spherical sampling schemes.
• Clustering (DBSCAN) and regression (SVR) for estimating coherent sound sources from power maps.
• find_sources wrapper function for automating estimation of source trajectories from an Ambisonic audio file.
• Functions to plot outputs.
• Implementations of Frame Recall and DOA Error performance metrics from DCASE 2019.

Requirements

• numpy
• soundfile
• resampy
• matplotlib
• scipy
• seaborn
• progressbar
• dipy
• scikit-learn

Usage

find_sources(input, *args, **kwargs)

input should be a path to an Ambisonic audio file.

*kwargs passed to sph_peaks_t:

• max_n_peaks=20 - the maximum number of peaks that will be saved per frame.
• audio_length_seconds=None - optional variable replacing output frame numbers with time in seconds.

*args passed to obj_trajectories:

• eps=0.1 - DBSCAN Eps parameter.
• min_samples=10 - DBSCAN MinPts parameter.
• relative_peak_threshold=0.5 - dipy rel_pk parameter.
• min_separation_angle=5 - dipy min_sep parameter.