Audio Data Augmentation

Audio augmentation applies transformations to training signals or their features to create varied examples. The goal is to make a model less sensitive to changes that should not alter the task label, such as background noise for some speech commands. Augmentation expands the effective variety of training inputs without claiming that synthetic clips replace real recordings. Noise mixing adds a separate signal at a chosen level. Room impulse response convolution approximates reverberation and room acoustics. Speed perturbation changes playback rate and usually shifts pitch as well; time stretching aims to change duration while preserving pitch, though algorithms introduce artifacts. Pitch shifting changes fundamental frequency while trying to preserve duration. Codec simulation reproduces distortions from compression or resampling. Spectrogram methods such as time and frequency masking hide portions of a representation during training. The label-preservation assumption is central. A small time stretch might retain a spoken word label, but an extreme stretch can make speech unintelligible. Pitch changes may alter speaker or emotion cues that matter to a task. Noise may mask the very acoustic event a classifier must detect. For sound-event classification, mixing two clips can require a multi-label target rather than copying one label. Transformations should reflect plausible deployment conditions and the target definition. Keep random training augmentation out of the primary evaluation split. Separately specified robustness tests or test-time augmentation need an explicit protocol; silently changing evaluation inputs obscures the comparison. Split recordings by speaker, source, or session before augmentation so transformed copies of the same original do not leak across partitions. Tune transformation ranges using domain knowledge and development data. Evaluate both clean and realistically corrupted inputs when both matter. Track the augmentation recipe and random seeds for reproducibility. Compare against unaugmented baselines and inspect failures, because augmentation can improve robustness to one shift while harming performance on another. More variation is not automatically better.