DeepFilterNet is hard-capped at 30 dB, with no stronger open model

(#4791) Feature Open: Under consideration voice

Current problem

DeepFilterNet is the strongest built-in suppressor, and loud noise still comes through. The limit is hard-coded at 30 dB: fluxer_app/src/features/voice/utils/noise_suppression/DeepFilter.ts DEEP_FILTER_ATTEN_LIM_DB = 30 Loud noise is turned down, not removed. Turning automatic gain control off makes the voice too quiet, which matches #3763. The choices in the current source are None, the browser's WebRTC suppressor, a noise gate, Speex, RNNoise, GTCRN and DeepFilterNet 3. None of those is stronger than DeepFilterNet.

Proposed change

Make that 30 dB cap a setting, and add a stronger open model beside DeepFilterNet. DPDFNet (Ceva, December 2025, Apache-2.0 code and models) is the open one with a claimed gain. On the authors' noisy multilingual test, PESQ is 3.14 for the small model and 3.20 for the large, against 2.76 for DeepFilterNet 3, at about 4x and 12x the compute. On a second test the gap is smaller (about 2.96 against 3.10 to 3.12). I found no independent comparison, and none against Krisp. I timed the 48 kHz streaming models on an i7-13700K, CPU only, on synthetic audio, so this is speed only: 1.1 ms per 10 ms of audio for the small model and 3.1 ms for the large. Both are comfortable on a desktop. A browser would be several times slower, and the 48 kHz models are about 7x (small) to 20x (large) the compute of DeepFilterNet 3, so it costs real time on a laptop or a phone. It still keeps anything that sounds like speech, so a TV or another person talking comes through, same as the others. Discord's suppressor is Krisp, which an open-source client cannot ship.

Platform

Linux, Web

Additional information

The 30 dB constant and the model list are from the current upstream source. The timing was on a desktop CPU, not in the browser.

Comments

Sign in with Fluxer to comment and vote.

No comments yet.