DeepFilterNet 3 vs RNNoise: a fan-noise test
Tested October 24, 2026 · RNNoise v0.2 (Xiph.Org, default model) · DeepFilterNet 3 as run by this site
Short answer: on our three fan-noise recordings both models removed the steady fan in the pauses, and automatic transcripts came out almost the same (0, 2 and 1 wrong words out of 92 with DeepFilterNet 3; 1, 3 and 1 with RNNoise). The audible difference is in how they do it: RNNoise switches the sound off completely for 13–21% of the recording, mostly in the pauses, while DeepFilterNet 3 leaves a very quiet, steady background. A 2025 study that tested both found DeepFilterNet 3 ahead on speech quality and intelligibility. RNNoise is about 80 times lighter, which is why it can run inside your browser.
Why it matters: free in-browser tools such as StayPut and Clyper run RNNoise on your own device, so nothing is uploaded. Server tools such as this site and VidClean run DeepFilterNet 3. We make audioenhancer, so we are not neutral: the samples and numbers are below, and you can repeat the test with the open-source programs listed under Method.
Listen: the same 20 seconds through both models
This is the desk-fan clip from our fan-noise page (the speaker's voice was pitch-shifted for privacy). Both results were set to the same loudness. Switch while it plays: playback continues from the same moment.
Listen for three things: the pauses (silent, or a soft steady background?), the start and end of each sentence (does the noise jump in and out?), and the voice itself (natural, or thin and watery?).
Our measurements on three fan recordings
Three one-take recordings of the same 92-word English script by the same speaker, 42 to 46 seconds each, all made near a running fan. Each was cleaned by both models, then raised back to the original loudness with the same rule this site uses, so neither result looks quieter just because it is softer overall.
| Recording | Quietest 10% of moments | Complete silence | Wrong words (of 92) |
|---|---|---|---|
| Take 1 · moderate fan | −40 → −91 / silent dBFS | 2% / 18% | 0 → 0 / 1 |
| Take 2 · loud fan | −26 → −65 / silent dBFS | 0% / 13% | 5 → 2 / 3 |
| Take 3 · loud fan | −28 → −80 / silent dBFS | 0% / 21% | 1 → 1 / 1 |
- Both remove the fan in the pauses. With RNNoise the quietest moments are digital silence, so this measure cannot rank the two; with DeepFilterNet 3 they sit 39 to 52 dB below the original noise.
- RNNoise gates. “Complete silence” is the share of 20-millisecond slices that come out as exactly zero. RNNoise does this for 13–21% of each recording, so the background drops to nothing between words and returns with the voice. Some listeners find that clean; others hear it as pumping.
- Transcripts are a tie. Automatic transcription (OpenAI Whisper, numbers written as words) differs by at most one word, which is within what a second run can change. A transcript counts recognisable words, not how natural the voice sounds.
- Loudness stayed within 0.6 dB of the original for both. Three recordings are not a benchmark: compare on your own file.
What a published study found
Maulana, Akbar and Iklima compared the two models on speech from the JL Corpus mixed with ESC-50 environmental noise (SINERGI 29(2), 2025). With each model's pre-trained weights:
| Model | STOI (intelligibility) | PESQ (quality) | SI-SDR (dB) | Processing time per second of audio |
|---|---|---|---|---|
| Before enhancement | 0.727 | 1.10 | −0.9 | — |
| DeepFilterNet 3 | 0.810 | 1.55 | 7.7 | 0.081 s |
| RNNoise | 0.743 | 1.23 | −10.6 | 0.001 s |
The study is small (five samples) and its noise is not fan noise, but it points the same way as the listening test: DeepFilterNet 3 keeps more of the voice, and RNNoise is far cheaper to run.
Which one to use
- Nothing may leave your computer, or the file is very long: an in-browser RNNoise tool such as StayPut or Clyper. Free, private, and fine for mild steady noise.
- A talk, tutorial or demo you will publish: try DeepFilterNet 3. This site cleans files up to 5 minutes free, copies the picture without re-encoding and gives you a 60-second preview of longer files.
- Whichever you use, compare the result with your original at the same volume before you replace it.
Try your own recording Compare 13 noise-removal tools
Questions
Is DeepFilterNet better than RNNoise?
On speech quality, usually yes: in a 2025 study DeepFilterNet 3 scored higher on intelligibility (STOI 0.81 vs 0.74) and quality (PESQ 1.55 vs 1.23). On our three fan recordings both removed the fan in the pauses and transcripts were within one word; RNNoise silenced 13–21% of each recording completely. RNNoise needs about 80 times less computing, so it can run in a browser.
Which noise-removal model do in-browser tools use?
StayPut and Clyper say on their pages that they run RNNoise on your device, so the file is not uploaded. This site and VidClean run DeepFilterNet 3 on their servers.
Can I run this test myself?
Yes. Both models are open source: build RNNoise v0.2 from Xiph.Org and run its demo program on 48 kHz mono audio, and run DeepFilterNet 3 with its deep-filter program. Use the same recording, match the loudness, and listen at the same volume.
Method
- RNNoise: Xiph.Org rnnoise v0.2 release, default built-in model, examples/rnnoise_demo on 48 kHz mono 16-bit audio.
- DeepFilterNet 3: the deep-filter program (v0.5.6) as run by this site, including its 6 dB headroom and length padding.
- Both: raised back to the original loudness (EBU R128), by up to 12 dB, with a limiter at −1 dBFS. The 20-second samples were set to the same loudness instead.
- Levels are from 20-millisecond slices of the whole file (16 kHz mono). Transcripts: OpenAI Whisper via fal.ai, October 24, 2026.
Sources
- RNNoise (Xiph.Org) and the DeepFilterNet repository
- M. I. Maulana, M. F. R. Akbar, Z. Iklima, “Performance of speech enhancement models in video conferences: DeepFilterNet3 and RNNoise,” SINERGI, vol. 29, no. 2, pp. 269–278, 2025
- H. Schröter et al., “DeepFilterNet: Perceptually Motivated Real-Time Speech Enhancement,” arXiv:2305.08227, 2023
- StayPut and Clyper noise-removal pages (RNNoise in the browser); VidClean (DeepFilterNet 3)
- Our fan recordings and measurements: fan-noise page