The WebRTC noise suppression estimates a running noise profile, so it removes
stationary noise but not keystrokes, which are transient and never look like
noise to it. Chiclet keyboards are as affected as mechanical ones. DeepFilterNet3
runs client-side as a LiveKit track processor and removes them.
Off by default and the model is only fetched once a user switches it on, so
nobody pays the download who does not want the filter. Default strength is 35
percent rather than full: measured, that already gives keystrokes gone with the
voice still natural, and more attenuation only adds artefact risk.
Assets ship with us instead of the package's default CDN, which would report
every participant's IP to a third party at call start and tie call setup to
foreign infrastructure. The Dockerfile gzips the model wasm, which the existing
top-level glob missed — 4.1 MB instead of 15.7 MB per client.
Browser noise suppression is switched off while the filter runs so the two do not
work against each other. Regulation goes through the model's own attenuation
limit, so there is no dry/wet mixer and no delay compensation to get wrong.
Decision and measurements: management ADR-0018 and issue #0054.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
