Work

Wav2Aug: A Drop-In Speech Augmentation Policy

Year: 2026
Type: dissertation
Abstract: SpecAugment is the dominant augmentation policy for automatic speech recognition (ASR), but it has been studied primarily for ASR and operates on spectrograms, limiting its applicability to other speech tasks and waveform-based models. We propose Wav2Aug, a drop-in waveform augmentation policy that randomly samples two augmentations from a calibrated pool of waveform transformations. Its parameter ranges are calibrated once and then reused across tasks and architectures, eliminating task-specific augmentation tuning. Experiments on ten speech benchmarks show that Wav2Aug achieves ASR performance comparable to SpecAugment while outperforming all reported SpecAugment comparisons on the evaluated non-ASR tasks, yielding an average relative error reduction of 12.8% over the strongest compatible SpecAugment comparison for each task. We further show that multi-view training provides additional gains, with two augmented Wav2Aug views offering the best trade-off between performance and computational cost. Kullback–Leibler (KL) consistency is also evaluated as an explicit agreement loss between views, where we show that its benefits are limited. Overall, Wav2Aug provides a single calibrated waveform policy that trans- fers across heterogeneous speech tasks, model families, and input representations. Despite keeping augmentation hyperparameters unchanged across tasks, Wav2Aug remains closely competitive with SpecAugment on ASR while improving over it on non-ASR tasks. Its parameter ranges, code, and installable Python library are released as open source.
Language: English
Open Access status: green
OpenAlex ID: W7221495843
Cites: 77
Cited by:
Related to: 10