A neural net that learned to shut up the background

What it does

RNNoise is a C library that strips noise from speech using a recurrent neural network. Feed it raw 16-bit mono PCM at 48 kHz and it spits back cleaner audio, also in raw PCM — no WAV headers, no hand-holding. It is built for real-time use, with AVX2 and SSE4.1 paths available to keep latency low.

The interesting bit

The project sits at an unusual intersection: it is a hybrid DSP/deep-learning approach from Xiph.Org, the same folks behind Opus and Vorbis. The model is small enough to quantize and embed directly as C data files, but you can also swap in your own trained weights at runtime. Training is fully documented and uses only public datasets — no proprietary audio collections required.

Key highlights

• Runs fully offline; no cloud APIs or network calls
• Supports custom model training with public speech and noise datasets
• Optional reverberation augmentation during training for room-aware models
• Provides both full and “little” (sparser) model variants
• Primary development happens on GitLab; GitHub is a convenience mirror

Caveats

• The API has a sharp edge: loaded model objects must outlive the RNNoise state, and you must not close the backing file while it is active
• Model format changed post-v0.1.1, so older external weights may not load without size-matching header tweaks

Verdict

Worth a look if you need embeddable, real-time speech denoising in a C codebase and do not mind wrestling with raw PCM. Skip it if you need plug-and-play WAV support or a polished Python wrapper — this is a library, not a product.