NR2 Noise Reduction

NR2 is a client-side spectral noise reduction method that complements the radio's built-in NR. It uses the Ephraim-Malah MMSE Log-Spectral Amplitude estimator, derived from Warren Pratt NR0V's WDSP library, and runs on your computer's CPU — not on the radio.

Contributed by @EI6JGB

For how NR2 compares with the other six client-side methods, see DSP Noise Mitigation.

Requirements

FFTW3 is required to build AetherSDR (the bundled WDSP library depends on it). Release builds include it. When building from source:

• Arch: sudo pacman -S fftw
• Ubuntu/Debian: sudo apt install libfftw3-dev
• macOS: brew install fftw
• Windows: run scripts/setup/setup-fftw.ps1

See Building from Source.

How it works

NR2 processes the demodulated receive audio on your CPU. It estimates the noise floor in each frequency bin and applies a spectral gain that suppresses noise while preserving speech. Like every client-side method, it processes the left and right channels independently, and it keeps its noise estimate across a transmission so it does not have to re-converge every time you unkey.

Using NR2

Turning NR2 on

• AetherRX: open Settings → AetherRX... (or the AetherRX button on the VFO flag's DSP tab), go to the AetherNR tab and select NR2 in the method row. NR2's settings appear below it.
• Shortcut: bind NR Cycle (Off/NR/NR2/NR4/DFNR) under Settings → Configure Shortcuts..., or to a MIDI control.

NR2 is a voice method: it switches itself off in CW, CWL, RTTY and DIGU/DIGL.

When to use NR2

• Steady noise — constant hiss or hum on a relatively quiet band, where you want to dial in exactly the right amount of suppression
• Weak signals where the radio's NR isn't enough
• Stacking with radio NR — NR and NR2 use different algorithms and can complement each other (enable NR on the radio, NR2 in AetherSDR)

NR2 is less effective against impulse noise (clicks, pops, lightning crashes); use the radio's NB or WNB for those.

FFTW wisdom (first use)

The first time you enable NR2, AetherSDR optimizes the FFT plans it uses (FFTW "wisdom"). An AetherSDR — FFTW Wisdom window appears while this runs:

• It shows progress, a Working activity indicator and an Elapsed time counter. Planning can take several minutes on some systems; if the progress pauses, the window says it is still working.
• The window is modeless — keep operating while it runs. On macOS it stays visible when AetherSDR is not the frontmost app.
• Clicking NR2 again while it runs brings the same window forward instead of starting a second run.
• Cancel stops after the FFT plan currently being computed. NR2 is then not enabled, and your audio is left unchanged.
• NR2 switches on when planning finishes, and the window closes itself.

The result is saved as soon as planning finishes, so a crash or force-quit afterwards does not lose it, and later launches skip the wait. The same cache also speeds up the WDSP receive setup on the Hermes-Lite 2.

The wisdom file is stored at:

Platform	Path
Linux / macOS	~/.config/AetherSDR/aethersdr_fftw_wisdom
Windows	%APPDATA%\AetherSDR\aethersdr_fftw_wisdom

To force a fresh optimization (for example after upgrading FFTW), quit AetherSDR, delete that file, and enable NR2 again.

Upgrading from an older version: older builds kept the file in a double-nested ~/.config/AetherSDR/AetherSDR/ folder. Current builds use the path above, so NR2 optimizes its plans once more after upgrading. You can delete the old file.

Reference

Settings

Setting	Range / options	What it does
Gain method	Linear, Log, Gamma, Trained	The speech model used to compute the suppression gain. Gamma is the default.
Noise estimation	OSMS, MMSE, NSTAT	How the noise floor is tracked. Try NSTAT if the noise changes quickly.
Reduction	0.50–2.00	How aggressively noise is suppressed.
Naturalness	0–0.15	Leaves a small residual noise floor, trading depth for a more natural sound.
Smoothing	0.50–0.98	Smooths the gain over time to reduce musical artifacts. Higher is smoother but slower.
Threshold	0.05–0.50	Speech-detector sensitivity. Lower keeps quiet voices but lets more noise through.
AE Filter (artifact elimination)	On / off	Reduces ringing and musical artifacts.
Noise fill (psychoacoustic)	Off by default	Mixes a little shaped noise back into the gaps between syllables. Fill level, Fill character and Fill bandwidth shape it.

Start with the defaults. See Troubleshooting if you hear musical warbling or weak voices disappear.

Known issues

• NR2 can sound scratchy or produce clicks and artefacts on voice (#632).

Troubleshooting

Musical warbling in the background

The gain is changing too quickly from frame to frame.

1. Raise Smoothing.
2. Turn on AE Filter (artifact elimination) if it is off.

Weak voices disappear

The speech detector is missing quiet speech and treating it as noise.

1. Lower Threshold.
2. If that is not enough, lower Reduction or raise Naturalness.

The FFTW Wisdom window takes a long time

AetherSDR is optimizing its FFT plans the first time you enable NR2. This can take several minutes on some systems.

1. Keep operating; the window is modeless.
2. Wait for it to finish. NR2 switches on and the window closes itself, and later launches skip the wait.
3. If you click Cancel, NR2 stays off; enable it again to restart planning.

NR2 turns off when you change mode

This is expected. NR2 is a voice method and switches off in CW, CWL, RTTY and DIGU/DIGL.

1. Switch back to a voice mode.
2. In CW, use the radio's NR and ANF instead.

NR2 does nothing against clicks and crashes

NR2 targets steady noise, not impulse noise.

1. Use the radio's NB or WNB for clicks, pops and lightning crashes.

See also

• DSP Noise Mitigation — how NR2 compares with the other client-side methods
• Aetherial Audio
• Building from Source
• Hermes-Lite 2