SmartSDR v4 – Noise Mitigation Quick Reference

Here is the very approximate ranges:

🔹 General Rule of Thumb

Most digital noise reduction systems in SmartSDR (and SDRs in general) perform best when the SNR of the desired signal is between about 0 dB and +20 dB.
Below 0 dB, there’s too little structure for the algorithms to reliably distinguish signal from noise.
Above +20 dB, the noise is already low enough that filtering offers little benefit and may start to color the audio unnecessarily.

Filter-by-Filter SNR Guidelines

NRL – Noise Reduction Leaky

• Optimal SNR Range: ≈ 0 dB → +15 dB
  
  
• Why: NRL relies on correlation detection — it can track structured content like speech or CW as long as there’s at least a small margin above the noise floor.
  
  
• Below 0 dB: The filter struggles to lock onto the desired signal and may flatten or “wash out” speech.
  
  
• Above +15 dB: Little improvement; may add mild “watery” artifacts.
  
  

ANFL – Adaptive Notch Filter Leaky

• Optimal SNR Range: Any SNR (works even with negative SNR)
  
  
• Why: It targets tonal interference, not random noise, so it doesn’t depend on SNR of the desired signal — only on the strength of the interfering tone relative to total signal power.
  
  
• Rule: The tone must be at least 10 dB stronger than the background to be recognized and notched effectively.
  
  

ANFT – Adaptive Notch Filter FFT

• Optimal SNR Range: Any, provided tone amplitude ≥ –110 dBFS and ≥ 10 dB above background.
  
  
• Why: FFT-based detection looks for narrow, stable spectral peaks. SNR of the main signal is largely irrelevant; what matters is tone prominence.
  
  
• Caution: In very low SNR conditions with fluctuating tones, detection latency may increase.
  
  

NRF – Noise Reduction Filter

• Optimal SNR Range: +3 dB → +20 dB
  
  
• Why: Spectral subtraction assumes the signal spectrum is distinct from the stationary noise floor.
  
  
• Below +3 dB: Noise and signal spectra overlap, causing audible “musical noise” or speech clipping.
  
  
• Above +20 dB: Filtering has negligible benefit.
  
  

NRS – Noise Reduction Speech

• Optimal SNR Range: 0 dB → +15 dB
  
  
• Why: The VAD component can detect speech reliably down to about 0 dB SNR; below that, it may misclassify weak speech as noise during syllable gaps.
  
  
• Benefit: Best audible quieting for average voice signals near the noise floor (3–10 dB SNR typical).
  
  

RNN – Recurrent Neural Network

• Operational SNR Range: –5 dB → +20 dB
  
  
• Why: The neural network is trained to recover speech even when buried below noise — it can function at slightly negative SNR because it models speech characteristics rather than statistical separation.
  
  
• Below –5 dB: Speech recovery becomes unreliable; noise dominates all features.
  
• Above +15 dB: Processing may subtly dull transients; otherwise transparent.
  

🔹 Summary Table

Takeaway

• For weak voice signals (SNR 0–10 dB): Try RNN, NRS, or NRL.
  
  
• For steady hums or carriers: Use ANFL or ANFT — they’re insensitive to speech SNR.
  
  
• For clean but slightly noisy audio (> 10 dB SNR): NRF provides subtle polish.
  
  
• For deep noise (negative SNR): Only RNN has any hope of recovering intelligible speech.