Anoise2

Anoise2 is an audio denoising algorithm and software framework that represents the second generation in the Anoise series. It is designed to suppress a wide range of noise types in speech and music signals while preserving perceptual quality and timbre. The system aims to perform well in real-world recording conditions, including mobile and conferencing scenarios.

Architecturally, Anoise2 uses a hybrid approach that combines time-frequency masking learned by neural networks with traditional spectral-subtraction techniques and Wiener filtering. It features an adaptive noise estimator that tracks non-stationary noise and a neural post-filter to reduce residual artifacts. The design supports streaming operation with configurable latency and can run on CPUs, GPUs, or embedded devices.

Key features include multi-channel support, optional dereverberation, and configurable noise profiles. It provides lightweight models for edge deployment and larger models for desktop use. The API is language-agnostic, with bindings for C++, Python, and common plugin formats compatible with digital audio workstations.

Applications span telecommunication and video conferencing, broadcasting, hearing aids, and music production, where reliable noise suppression, artifact control, and real-time performance are valued.

Evaluation and reception: Anoise2 is validated on standard datasets using metrics such as PESQ, STOI, and SI-SDR, and is benchmarked against traditional baselines as well as modern neural methods. Limitations include possible speech distortion or musical artifact under highly dynamic noises, and performance depends on dataset coverage and model configuration.

Availability: Anoise2 is released as an open-source project under a permissive license, with a reference implementation, documentation, and example pipelines hosted in a public repository.