audiomentations

Why we included this project

Teams training audio models on a single clean dataset often find those models fall apart on real-world recordings, which carry background chatter, room acoustics, or compression artifacts. Audiomentations closes that gap with a wide set of ready-made perturbations: noise injection, pitch and time shifting, impulse responses, EQ filters, gain transitions, and mp3 compression. Each transform takes its own probability, so you can assemble a randomized augmentation pipeline instead of hand-tuning every step. The API follows the familiar albumentations style, so anyone who has done image augmentation can put together a working Compose chain in minutes. It runs on CPU, handles mono and multichannel audio, and plugs straight into TensorFlow/Keras or PyTorch training loops, which makes it a practical choice for speech recognition, sound classification, and music work. PyTorch users who need GPU-accelerated transforms can turn to the companion torch-audiomentations library next.