Choosing the right noise reduction: NR2, NR4, DFNR, MNR¶
AetherSDR provides four client-side noise-reduction engines. This page describes what each engine does, when to use it, and where to find its controls so you can pick the right one for your operating conditions.
Before you start¶
- Open AetherDSP Settings via
Settings > AetherDSP Settings.... - The NR engine you configure here is client-side only; it does not require a radio connection.
Steps¶
- Go to
Settings > AetherDSP Settings.... - Click the tab for the engine you want to use: NR2, NR4, DFNR, or MNR.
- Adjust the controls on that tab (see the table below).
- Close the dialog. Settings are saved automatically.
What each control does¶
NR2 — musical-noise reduction¶
A frequency-domain noise reducer designed to minimise the tonal "birdie" artefacts common in spectral subtraction. Good starting choice for SSB voice in moderate QRN.
| Control | Kind | Default | Range | Setting key |
|---|---|---|---|---|
| Gain Method | Radio buttons | Gamma | Linear | Log | Gamma | Trained | NR2GainMethod |
| NPE Method | Radio buttons | OSMS | OSMS | MMSE | NSTAT | NR2NpeMethod |
| AE Filter (artifact elimination) | Checkbox | Enabled | — | NR2AeFilter |
| Reduction Depth: | Slider | 1.50 | 0.50–2.00 | NR2GainMax |
| Smoothing: | Slider | 0.85 | 0.50–0.98 | NR2GainSmooth |
| Voice Threshold: | Slider | 0.20 | 0.05–0.50 | NR2Qspp |
| Reset Defaults | Button | — | — | — |
Gain Method selects how NR2 maps noise estimates to gain reduction. Gamma matches typical speech amplitude patterns and is the default. Trained uses a model built from real speech and noise samples. Linear and Log trade perceptual accuracy for simpler computation.
NPE Method selects the noise power estimator. OSMS (Optimal Smoothing Minimum Statistics) tracks the noise floor using a running minimum and suits slowly varying noise. MMSE minimises expected estimation error. NSTAT adapts to noise that changes rapidly over time.
AE Filter (artifact elimination) applies a post-filter to reduce ringing and musical artefacts. Leave it enabled unless you are experimenting with very low Reduction Depth values.
Reduction Depth: controls maximum suppression. Higher values remove more noise but risk speech distortion. 1.50 is the default.
Smoothing: controls how quickly the noise estimate follows changes. Higher values are steadier but slower to adapt.
Voice Threshold: is the speech-presence-probability threshold. Lower values protect quiet speech but may allow more noise through.
Reset Defaults restores: Gamma / OSMS / AE Filter on / 1.50 / 0.85 / 0.20.
NR4 — libspecbleach¶
A separate spectral-bleaching engine with its own noise estimator and additional shaping controls. Useful when NR2 leaves residual noise or when you want dB-calibrated reduction targets.
| Control | Kind | Default | Range | Setting key |
|---|---|---|---|---|
| Noise Estimation Method | Radio buttons | SPP-MMSE | SPP-MMSE | Brandt | Martin | NR4NoiseEstimationMethod |
| Adaptive Noise Estimation | Checkbox | Enabled | — | NR4AdaptiveNoise |
| Reduction (dB): | Slider | 10.0 dB | 0.0–40.0 dB | NR4ReductionAmount |
| Smoothing (%): | Slider | 0 | 0–100 | NR4SmoothingFactor |
| Whitening (%): | Slider | 0 | 0–100 | NR4WhiteningFactor |
| Masking Depth: | Slider | 0.50 | 0.00–1.00 | NR4MaskingDepth |
| Suppression: | Slider | 0.50 | 0.00–1.00 | NR4SuppressionStrength |
| Reset Defaults | Button | — | — | — |
Noise Estimation Method — SPP-MMSE balances noise estimation with speech preservation. Brandt uses recursive smoothing over critical frequency bands and suits non-stationary noise. Martin uses running spectral minima and is robust for slowly varying noise floors.
Adaptive Noise Estimation enables continuous re-estimation of the noise floor. Disable it only if the noise environment is static and you want a fixed floor.
Reduction (dB): sets the maximum reduction in dB. Start at 10 dB and increase if noise remains.
Smoothing (%): applies time-domain smoothing to the noise estimate.
Whitening (%): flattens the residual noise spectral shape after reduction.
Masking Depth: controls the depth of spectral masking applied.
Suppression: sets the overall suppression strength. Higher values are more aggressive.
Reset Defaults restores: SPP-MMSE / Adaptive on / 10.0 dB / 0 / 0 / 0.50 / 0.50.
DFNR — DeepFilterNet3¶
A neural-network-based noise filter. Suited for strong broadband noise where conventional spectral methods fall short. Has the highest CPU cost of the four engines.
| Control | Kind | Default | Range | Setting key |
|---|---|---|---|---|
| Attenuation Limit | Slider | 100 dB | 0–100 dB | DfnrAttenLimit |
| Post-Filter Beta | Slider | 0.00 | 0.00–0.30 | DfnrPostFilterBeta |
Attenuation Limit sets the maximum noise attenuation DeepFilterNet3 will apply. 0 is passthrough; 100 is maximum attenuation. Reduce this value if the neural filter over-suppresses weak signals.
Post-Filter Beta adds an extra suppression stage on top of the neural filter output. Leave at 0.00 unless residual noise remains after adjusting Attenuation Limit.
MNR — macOS only¶
An MMSE-Wiener noise reducer with asymmetric gain smoothing, available only on macOS.
| Control | Kind | Default | Range | Setting key |
|---|---|---|---|---|
| Enable MNR (macOS only) | Checkbox | (read from audio engine) | — | MnrEnabled |
| Strength | Slider | 100 | 0–100 | MnrStrength |
Enable MNR (macOS only) turns the engine on or off. The initial state reflects the current audio engine state.
Strength sets aggressiveness. 0 is the mildest; 100 is maximum. Persisted internally as a normalised value of 0.00–1.00.
MNR is not available on Linux or Windows. The MNR tab is still present but Enable MNR (macOS only) will have no effect on non-macOS systems.
RN2 — RNNoise¶
The RN2 tab is informational only. RNNoise has no adjustable parameters in AetherDSP Settings. Enable or disable the engine from the overlay menu.
BNR — NVIDIA¶
The BNR tab is informational only. BNR intensity is controlled from the overlay menu, not from AetherDSP Settings.
Tips¶
- Run only one noise-reduction engine at a time. Chaining multiple engines can cause speech artefacts and adds CPU load.
- For SSB voice with moderate band noise, start with NR2 at its defaults before trying NR4 or DFNR.
- If you are on macOS and prefer a lighter CPU load, MNR is the lowest-overhead option.
- DFNR's Attenuation Limit at 100 dB can suppress very weak signals along with the noise. Reduce it to 40–60 dB on marginal paths.
- On the NR2 tab, if speech sounds hollow or "underwater", lower Reduction Depth: toward 0.80–1.00 or switch Gain Method from Gamma to Log.
- Use Reset Defaults on the NR2 or NR4 tab to recover a known-good starting point after experimental changes.
Troubleshooting¶
- Speech sounds hollow or musical artefacts are audible on NR2 — Reduce Reduction Depth: or confirm AE Filter (artifact elimination) is enabled.
- NR4 is not reducing noise enough — Increase Reduction (dB): and enable Adaptive Noise Estimation if it is off.
- DFNR removes weak signals along with the noise — Lower Attenuation Limit from 100 toward 40–60 dB.
- MNR tab is present but has no effect — MNR is macOS-only. On Linux or Windows, use NR2, NR4, or DFNR instead.
- NR2 or NR4 settings did not persist after restart — Settings are saved automatically on each control change. If values revert, click Reset Defaults and re-enter the desired values to force a save.
Related¶
- AetherDSP Settings overview
- Tune NR2 reduction depth and voice threshold
- Switch NR2 gain method between Linear, Log, Gamma and Trained
- Change NR2 noise power estimator (OSMS/MMSE/NSTAT)
- [Adjust NR4 reduction amount in dB](../../features/aether-dsp/adjust-