Audio AI GUIDE

Speech Denoising with RNNoise

RNNoise is a tiny, fast neural network that strips background noise from speech in real time.

2 min readLast updated

Overview

Created by Xiph.Org's Jean-Marc Valin, it pairs classic signal processing with a small recurrent network so it runs on ordinary CPUs and even embedded devices.

Deep Dive

RNNoise, released in 2017, was designed for low-latency noise suppression in voice calls. Rather than learning everything end-to-end, it splits speech into about 22 frequency bands modeled on the human ear (a Bark-like scale) and uses a recurrent neural network with Gated Recurrent Units to estimate a gain (0 to 1) for each band per frame. Those gains attenuate noisy bands while keeping speech-dominated bands intact. A complementary pitch filter cleans up residual noise between harmonics of voiced speech. The whole model has roughly 85,000 weights, runs faster than real time on a single CPU core, and is open source under a BSD license, which is why it was integrated into projects like the Opus codec ecosystem, Mumble, and OBS Studio.

Technical Insight

The key design choice is operating on perceptual band gains instead of raw spectral bins. By predicting only ~22 gain values per frame, the GRU network stays tiny and avoids musical-noise artifacts common in older spectral-subtraction methods. Hand-crafted features (band energies, pitch period, pitch correlation) feed the network, blending DSP knowledge with learning. A separate voice-activity output helps gate gains during pure-noise frames.

Strategic Impact

Access and reach

It improves accessibility through transcription, narration, and voice interfaces.

Cost and budget

Media teams can ship polished audio faster with smaller budgets.

Speed and scale

Customer-facing systems can process spoken interactions at larger scale.

The Future of Speech Denoising with RNNoise

RNNoise inspired a wave of lightweight real-time enhancement work; its successor research (PercepNet, DeepFilterNet) pushes quality higher while keeping CPU budgets small. Expect denoisers to embed directly into headsets, hearing aids, and conferencing chips, to combine with echo cancellation and dereverberation, and to use perceptual and even generative objectives. The hybrid DSP-plus-small-network recipe remains influential wherever low latency, low power, and open-source licensing matter more than raw model size.

Real-World Implementation

Suppressing keyboard clatter and fan hum during video calls in apps that bundle RNNoise.

Cleaning up a streamer's microphone in OBS Studio via the built-in RNNoise noise-suppression filter.

Improving intelligibility of voice chat in games and VoIP tools like Mumble on low-power hardware.

Preprocessing noisy field recordings so downstream speech recognition gets a cleaner signal.

Risks & Guardrails

Voice misuse and impersonation risks increase when consent is missing.

Accuracy can drop across accents, dialects, or noisy environments.

Synthetic audio can be mistaken for authentic speech without clear labeling.

Implementation Roadmap

1

Obtain explicit consent for voice capture, cloning, and reuse.

2

Test quality across diverse speakers and background conditions.

3

Define when a human must review or approve outputs.

4

Label synthetic audio and keep provenance records for accountability.

Keep Exploring

Free newsletter

Keep up with AI in 3 minutes a day

One short email each weekday with the three AI stories that actually matter. Free forever, no ads.

One email each weekday. Unsubscribe in one click. We never sell or share your address.

Test yourself

Take the Speech Denoising with RNNoise quiz

Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.

Start quiz

Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation

Next guide

Speech Separation and the Cocktail Party Problem

Frequently asked questions

What is Speech Denoising with RNNoise?

RNNoise is a tiny, fast neural network that strips background noise from speech in real time. Created by Xiph.Org's Jean-Marc Valin, it pairs classic signal processing with a small recurrent network so it runs on ordinary CPUs and even embedded devices.

What does RNNoise primarily predict for each short frame of audio?

RNNoise estimates per-band gains that attenuate noise-dominated bands while preserving speech-dominated ones, rather than working on every spectral bin.

What type of neural network is at the core of RNNoise?

RNNoise uses a compact recurrent neural network built with Gated Recurrent Units, giving it temporal memory while staying tiny.

Why does RNNoise use perceptual frequency bands instead of raw spectral bins?

Predicting only ~22 band gains keeps the network small, fast, and less prone to the musical-noise artifacts that plague per-bin methods.

Roughly how large is the RNNoise model?

RNNoise has on the order of 85,000 weights, small enough to run faster than real time on a single CPU core.

What is the role of the pitch filter in RNNoise?

A comb/pitch filter exploits the harmonic structure of voiced speech to suppress noise sitting between the harmonics, complementing the band gains.