Noise Suppression

General definition

Noise Suppression is the processing step that removes or reduces unwanted background sound from a microphone signal before it is encoded and sent, so the person on the other end hears speech rather than the room. It sits in the audio pipeline of every calling product next to echo cancellation and automatic gain control, and it runs on the sending side, where the raw signal is available.

Classic noise suppression works on the spectrum of the signal. It estimates the noise floor during pauses in speech, then attenuates the frequency bands where noise dominates; spectral subtraction and Wiener filtering are the standard techniques. This handles stationary noise well: fans, air conditioning, engine hum. It struggles with non-stationary sounds such as keyboard clicks, a barking dog or a second voice, and aggressive settings produce the hollow, underwater artefact known as musical noise.

AI-based suppression trains a neural network to separate speech from everything else. RNNoise, released in 2017 by the Xiph.Org Foundation, showed that a small recurrent network could run in real time on an ordinary CPU, and commercial models from Krisp, NVIDIA, Discord, Zoom and others followed. These models remove non-stationary noise that spectral methods cannot, at the cost of more compute and, occasionally, cutting off quiet speech or music the user wanted to send.

• In WebRTC, the audio processing module ships a built-in suppressor that browsers enable by default; the getUserMedia constraint noiseSuppression turns it on or off per track
• Suppression is applied before the codec, so it also lowers bitrate: a quieter signal compresses better with Opus
• It does not fix jitter or packet loss; those are network problems handled by the jitter buffer and packet loss concealment
• For music, podcasts or clinical listening, turn it off: what it removes as noise may be the content

In the Ethora ecosystem

Ethora calls are built on WebRTC, so the browser and mobile WebRTC stacks provide the standard audio processing chain, including noise suppression, echo cancellation and gain control, on every participant’s device before audio reaches the SFU. Because the SFU forwards media rather than decoding it, cleaning up audio is a client-side job, which is also where it belongs: the raw microphone signal never leaves the device.

The same design keeps your options open. A product that needs stronger suppression can add an AI model in the client before the audio track is handed to WebRTC, without any change to the server, and a product where fidelity matters more than silence can leave suppression off in the client. Audio-only calls, available as an opt-in in the voice and video call API, benefit the most, since voice is the whole experience.