What is zipenhancer?

This is a Python service that removes background noise from audio files. It takes the ZipEnhancer noise-reduction model, originally released by Alibaba's research lab via a platform called ModelScope, and reimplements the inference code in plain PyTorch so it can run without the ModelScope pipeline wrapper. The result is wrapped in a FastAPI web service that accepts audio file uploads and returns cleaned audio. The service supports three noise-reduction models switchable per request: ZipEnhancer for lightweight use, FRCRN for real-time scenarios, and MossFormer2 for higher quality output. It handles audio of any length by processing the file in four-second sliding windows with overlap, which prevents running out of GPU memory on long recordings. Stereo input produces stereo output, and the original sample rate and bit depth are preserved by default. Output can be saved as WAV, FLAC, MP3, or OGG. The API has three main endpoints: a health check, a model listing endpoint, and a denoise endpoint that accepts either a single audio file or a folder of files for batch processing. Parameters control the output format, bit rate, sample rate, and whether to apply volume normalization. GPU acceleration is supported on NVIDIA graphics cards and can reach processing speeds more than twenty times faster than real time on high-end hardware. Setup involves creating a Python 3.10 environment, installing dependencies, copying an example environment file, and starting the service with a single command. The readme and documentation are written in Chinese. The project is released under the MIT license and is aimed at developers who want to add noise reduction to audio pipelines, meeting recordings, or audio preprocessing workflows without relying on a vendor-managed model platform.