iver56/torch-audiomentations

Transforms are implemented as `nn.Module` subclasses and support GPU acceleration (CUDA), making them composable directly into neural network architectures with minimal training overhead. The library provides flexible randomization control through `per_batch`, `per_example`, and `per_channel` modes, allowing fine-grained augmentation strategies; most transforms are differentiable to enable end-to-end training. It includes 15+ waveform transforms (pitch shift, filtering, gain modulation, impulse response convolution) and handles batched multichannel audio natively.

1,136 stars.