This speech denoiser borrows its brain from music separation

What it does

Facebook Research’s denoiser is a causal speech-enhancement model that runs in real time on a laptop CPU. It uses an encoder-decoder architecture with skip-connections—the Demucs network originally designed for music source separation—to remove stationary and non-stationary background noise and room reverb. The system is optimized jointly in the time and frequency domains using multiple loss functions, and it can pipe cleaned audio into video-conference software via a loopback interface.

The interesting bit

The model works directly on the raw waveform, and it adapts the Demucs music-separation architecture for speech—a lateral move that is unusual in the audio-processing space.

Key highlights

• Runs fully causal and real-time on a standard laptop CPU (tested on a 2 GHz quad-core Intel i5 with DDR4 memory).
• Removes stationary and non-stationary noise plus room reverberation.
• Ships with pretrained checkpoints (dns48, dns64, master64) and supports fine-tuning.
• Can be used as a live audio filter for video calls on macOS and Linux via loopback interfaces.
• Applies data augmentation directly on the raw waveform during training.

Caveats

• Audio can become “crunchy” or distorted if the machine cannot process frames fast enough; the README explicitly warns that older DDR3 memory may struggle.
• No official support for Windows or platforms beyond macOS and Linux.
• Very high noise levels can cause the model to introduce artifacts rather than clean speech.

Verdict

Worth a look if you need a low-latency, CPU-friendly speech enhancer for live conferencing or embedded pipelines. Skip it if you need guaranteed production polish on Windows or heavily overloaded hardware.