Table 3 Description of selected data augmentation techniques commonly used in SELD

SpecAugment¹⁰²

Randomly masks blocks along the time and/or frequency dimensions of a spectrogram, akin to time-frequency masking or cutout¹⁰³, to improve model robustness against partial data loss.

ACS⁷⁴

Applies directional transformations to multi-channel recordings (FOA or MIC) by swapping or rotating channels, effectively simulating up to eight different DOA configurations while preserving reverberation characteristics.

MCS⁷⁴

Synthesizes new multi-channel audio by combining beamformed spectral features with spatial cues derived from covariance analysis, simulating novel spatial audio scenarios.

Mixup¹⁰⁹

Blends two monophonic audio waveforms and their labels. Generally used to create synthetic polyphonic mixtures to improve robustness in densely polyphonic acoustic scenes.

FilterAugment¹⁰⁴

Simulates spectral coloration imparted by diverse acoustic environments by applying random gain offsets to randomly sized frequency bands.

Frequency Shifting⁵¹

Reflects (shifts) a randomly selected frequency band upward or downward, altering the spectral distribution while largely preserving the overall structure of the spectrogram.

SpatialMixup¹⁰⁶

Applies selective directional gains to multi-channel audio signals to simulate variations in directional loudness and suppression, increasing the spatial diversity of the data.

SpecMix¹⁰¹

Mixes two spectral feature representations (e.g., spectrograms) to form a new mixture, aiming to retain crucial frequency content from both sources.