johnjaejunlee95/deep-anc-reproduced

Deep ANC reproduced is a PyTorch reimplementation of a research paper on active noise control, a technique used to cancel out unwanted sound using deep learning. The author is clear that this is not the official version of the paper's code. It is a personal reproduction, so results and details may differ from what the original researchers published, and some patterns were borrowed from other public active noise control projects. The code is organized so a reader can find their way around fairly easily. There is a main script that handles both training and inference, a folder with the neural network model definition, a folder with helper utilities for things like audio processing and checkpoints, a folder holding filter files needed for the noise control math, sample audio files for testing, and a set of regression tests. Setting it up involves creating a conda environment with Python 3.10, then installing PyTorch built for CUDA 11.8, either through conda or pip, followed by the rest of the dependencies from a requirements file. This means a GPU with CUDA support is expected for realistic use, though a CPU option likely still works for smaller tests. Training requires a separate synthetic dataset, linked from Google Drive, which itself is based on data from a different published noise control project. The dataset needs specific CSV files marking training, validation, and test splits, each pointing to file paths within the dataset. Once training finishes, the best and most recent model checkpoints are saved automatically. After training, the project can run inference on sample noise recordings and produce output audio files, using the same filter data as training. There is also a simple test suite using pytest to confirm the code still runs correctly, alongside a compile check across the main script and helper files.