● For the Icom IC-7300 · free for licensed amateurs
Noise Reduction sits between your IC-7300's USB audio and your speakers and cleans it up live. It takes out band hiss, static crashes and carriers, so the weak voice under them is easier to copy. It has three AI engines, a look-ahead noise blanker and a CW filter that locks on to a signal you can barely hear.
Fig. 1 · The Weak SSB demo on AI Engine 3, 15.3 dB of noise removed at that moment.
This is the app's own waterfall from the Weak SSB demo, which has fading and band noise. On the left is the radio's raw audio; on the right is what you hear with AI Engine 3. The noise floor drops away and the speech stands out clearly.
RAW RADIO AUDIO
WHAT YOU HEAR
Fig. 2 · Screen captures from the app's built-in Weak SSB demo (AI Engine 3, 70 % strength, NB Low). The two halves were captured on two passes of the same demo recording.
The engines are the core of it, and the app adds the tools an operator needs around them. Everything runs in real time on your PC, and it keeps running when the window is minimised.
Classic, plus AI Engines 1, 2 and 3. Switch with one key while you listen, and keep the one that suits the band and the signal.
The app estimates the signal-to-noise ratio as you listen. As a station gets stronger it gradually blends back to the original audio, so strong signals still sound natural, not processed.
Static crashes and impulse noise are removed before they reach the engines. It looks ahead in the audio, so it catches the start of each crash. Three levels: Low, Med and High.
Tuning carriers and heterodynes are found and notched out automatically. The voice underneath stays intact.
Clarity emphasises the part of the voice that carries the words. Leveler smooths out QSB fading, so a fading station stays at a steady volume.
TRACK searches between your low and high cut, locks on to the CW signal and follows it with a 50–400 Hz filter and a keying gate. You hear it at the pitch you choose.
The live spectrum shows the radio audio, what you hear and the noise floor. The waterfall switches between Before and After, and a meter shows how many dB of noise are being removed.
Weak SSB, Static, Weak CW and Strong are ready to use with one click. Save your own settings for each band or contest.
Built-in demos show what each setting does before you connect the radio. Open file… lets you process a recording too.
Each engine has its strengths. The neural engines are best when a signal is buried; the statistical ones are lightest on strong, clean signals. AI Engine 3 is the most powerful of them.
The original algorithm. It tracks the noise floor with minimum statistics and applies a Wiener gain. Kept for comparison, and it scores best of all on strong signals.
Speech-presence noise tracking with a log-spectral gain. It's a statistical estimator, not a neural network, and it has the lowest delay of the AI engines.
A small recurrent neural network. It handles static crashes well and barely uses any CPU.
A larger deep neural network that filters the noise frequency by frequency. It's the strongest engine on weak, noisy signals; it adds about 40 ms of delay.
Hybrid mode. When AI Engine 2 or 3 is selected, AI Engine 1 also runs after it at half strength to clean up what is left. With AUTO lit, the app fades the AI processing back as the signal-to-noise ratio rises from about 8 dB to 20 dB. That stops the neural engines, which learned on microphone speech, from colouring strong SSB.
48 kHz throughout. In CW with TRACK on, the tracking filter and keying gate take the place of the engines.
The layout follows the front panel of a transceiver: scope and waterfall on the left, keys and knobs on the right. Everything fits on a 1366 × 768 laptop screen without scrolling.
Each engine was benchmarked offline on a synthetic SSB contact: three voices, a 300–2700 Hz audio band, fading, band noise and static crashes. Scores are STOI, a standard measure of speech intelligibility from 0 to 1, where higher is better. Each engine ran at 60 % strength with the noise blanker on Low.
| Engine · SNR | 25 dB | 10 dB | 5 dB | 0 dB | −6 dB | 0 dB + static |
|---|---|---|---|---|---|---|
| Raw audio | 0.83 | 0.70 | 0.64 | 0.57 | 0.48 | 0.46 |
| Classic | 0.91 | 0.75 | 0.66 | 0.57 | 0.48 | 0.38 |
| AI Engine 1 | 0.90 | 0.76 | 0.69 | 0.60 | 0.49 | 0.50 |
| AI Engine 2 | 0.89 | 0.76 | 0.69 | 0.60 | 0.50 | 0.55 |
| AI Engine 3 | 0.89 | 0.76 | 0.70 | 0.62 | 0.51 | 0.55 |
STOI speech intelligibility: about 20 % better than the raw audio.
How closely the audio follows the real keying: more than three times better.
Download the installer, which needs no admin rights, or the portable exe. Everything, including the AI models, is inside.
On the IC-7300, set ACC/USB Output Select to AF and connect the USB cable.
In Settings › Privacy & security › Microphone, turn on Let desktop apps access your microphone. Windows counts the radio's USB audio as a microphone.
Press Start radio, register once (it's free) and pick a preset. Press B at any time to hear the audio without processing.
Windows 10 or 11, 64-bit. The demos work straight away; live radio audio needs a free key from VU3RAZ.
The app isn't code-signed yet, so Windows SmartScreen may say "Windows protected your PC" the first time. Click More info → Run anyway. You can check each file against the SHA-256 shown above.
Noise Reduction is free for licensed radio amateurs. You register your callsign with VU3RAZ and get a key that unlocks live radio audio on that PC.
Reinstalling or updating the app keeps your registration. A different PC, or the same PC after Windows is reinstalled, has a new machine code and needs a new key.
AI Engines 2 and 3 are neural networks, built on RNNoise and DeepFilterNet3. AI Engine 1 is a statistical speech-presence estimator (OM-LSA) and contains no neural network. All of them run on your own PC; no audio leaves it.
No. Every model, including AI Engine 3, is built into the app. Once installed, it works with no connection at all.
AI Engine 3 adds about 40 ms, and AI Engine 1 has the lowest delay. The app shows the total delay in its top bar, so you can choose what suits your operating.
Light the AUTO key. The app then fades the AI processing back as signals get stronger. The Strong preset (AI Engine 1, light strength) is also made for loud stations and contests.
It was built and tested for the IC-7300. It processes any audio input Windows can see, so other rigs with USB audio or a sound-card interface should work too, but they haven't been tested.
That warning appears because the app isn't code-signed yet. Click More info → Run anyway. You can check that your download matches the SHA-256 published on this page.
Noise Reduction