RNNoise and DeepFilterNet: What to Test on a Board

The RNNoise and DeepFilterNet sources we checked publish no microcontroller-class cost, so measure it on your board.

The RNNoise paper aims for real-time operation at 48 kHz on a low-power processor. DeepFilterNet2 reports a real-time factor of 0.04 on a notebook Core-i5. One preprint found original audio had lower semantic word error than enhanced audio in all 40 tested configurations.

If you are choosing a denoise stage for a voice device, two questions come first. Does a neural model fit the chip, and does it help or hurt the speech recognizer downstream?

This article collects what the RNNoise and DeepFilterNet sources actually say, and lists what to measure on your own board and what to ask before you choose. All source facts are from public papers and repository pages as of October 31, 2026.

This is an engineering summary for general audio products. Where a vendor or author has not published a number, we say so and leave the measurement to you.

What Do the Published Sources Say About RNNoise and DeepFilterNet?

Only five facts below are published. Each row links the page it comes from, as of October 31, 2026.

What Can Go Wrong?

These points are engineering reasoning from the facts above. They describe no company's incident, and we have no measured results to report.

• The 48 kHz orientation. RNNoise's README and the DeepFilterNet repo both describe 48 kHz audio. A 16 kHz capture path needs resampling or a different model, and each extra stage adds delay and processing for you to measure.
• A desktop speed number says nothing about an MCU. An RTF of 0.04 on a Core-i5 says the model is light on that machine. The paper gives no cycles per frame on a microcontroller.
• Latency has no published figure here. The DeepFilterNet repo page gives no latency figure. If your product has a spoken-turn budget, the denoiser's lookahead counts toward it and has to be measured on the whole path.
• Cleaner audio for people can be worse input for a recognizer. In the one study we cite, original noisy audio gave lower semantic word error than enhanced audio in all 40 configurations. That is one study, so it flags a risk without proving it for your stack.
• Quality claims need your own noise. Published results use the authors' test sets. Your enclosure, microphone placement and wind or motion noise decide what the denoiser sees.

What Should You Measure on Your Own Board?

The sources above set no pass line for your product. We have not run these tests for this article and this page reports no results.

What Should You Decide Before Choosing?

Take this list to your audio firmware lead, your recognizer vendor or whoever builds the pipeline.

• Who consumes the cleaned audio: people, a recognizer, or both?
• Does the recognizer already include its own noise handling?
• What is the capture sample rate, and what does the model need?
• Where does resampling happen and who owns it?
• What CPU, RAM, flash and power budget does the denoiser get after wake word, codec, radio and application code?
• Do you need the neural model, or does a spectral baseline meet your own measured requirement?
• What is the latency budget, and how much of it can the lookahead use?
• Can the denoiser be switched off per use case, such as recognizer input versus human playback?
• Which noise conditions define success for your product, and are they in your test set?
• What are the license terms of the model and training data, and who checks them?

Questions About Noise Suppression on Small Chips

Does a neural denoiser always improve speech recognition?

No published source we checked says so. One arXiv preprint reported that the original noisy audio gave a lower semantic word error rate than enhanced audio in all 40 tested configurations. That is one study. Test your own ASR path with the denoiser on and off. arXiv 2512.17562, accessed 2026-10-31.

Can I run RNNoise on a 16 kHz voice pipeline?

The RNNoise repo README describes raw 16-bit mono PCM at 48 kHz, and the DeepFilterNet repo says only 48 kHz wav files are currently supported. A 16 kHz pipeline therefore needs resampling around the denoiser or a different model. Both choices add work and latency that you have to measure. RNNoise README, DeepFilterNet repo.

What does a real-time factor of 0.04 tell me about a microcontroller?

Very little. DeepFilterNet2 reports 0.04 on a notebook Core-i5 CPU, which means processing takes 4% of the audio duration on that machine. The paper does not give a microcontroller number, so any conversion to an MCU clock is an optimistic lower bound at best, because your core lacks the host's vector units. Measure on your board. DeepFilterNet2 paper.

Is there a published latency figure for DeepFilterNet?

No figure appears on the DeepFilterNet repo page we read, so measure end-to-end latency on your own signal path. DeepFilterNet repo, accessed 2026-10-31.

For the product category, see audio and voice hardware development.

Primary Sources

All pages below were accessed on 2026-10-31. Repositories and preprints change, so confirm figures on the live pages.

• RNNoise paper (arxiv.org/abs/1709.08243), accessed 2026-10-31.
• RNNoise README (raw.githubusercontent.com/xiph/rnnoise/main/README), accessed 2026-10-31.
• DeepFilterNet2 paper (arxiv.org/abs/2205.05474), accessed 2026-10-31.
• DeepFilterNet repository (github.com/Rikorose/DeepFilterNet), accessed 2026-10-31.
• Enhancement preprocessing and ASR study (arxiv.org/abs/2512.17562), accessed 2026-10-31.

Originally published on Duobay, where we write about building connected devices: electronics, firmware, enclosure and certification.