Guide Neural noise remover

Guide: the neural noise remover

What it is. An optional filter that removes steady background noise from speech. It uses RNNoise (Xiph.Org, BSD-3-Clause), a small neural network built into OnAir. It is off by default and it only touches the audio you hear and record; it does not change what the decoders see.

Where to find it

• General receiver: the checkbox Noise remover (neural) under the audio filters. It is offered for NFM, WFM and AM only. The setting and the strength are remembered per demodulator, so switching to USB, LSB or CW never has it on.
• Airband: the same checkbox in the squelch row, working on the mixed audio of all channels.

It is not offered for SSB, CW, music or data modes. RNNoise is trained on speech and treats anything else as noise, so music would be damaged. The label next to the box says it: "Speech only; do not use on music or data modes."

Step by step

1. Tune a voice signal: an FM two-way channel in the General receiver (NFM) or an airband channel.
2. Tick Noise remover (neural).
3. A strength slider appears, 0 to 100 %. It blends the original (0) and the cleaned audio (100). RNNoise itself has no strength control; the blend is OnAir's.
4. Listen while you move the slider: right removes more noise, left keeps more of the original. Full strength can sound thin on weak signals, so try a lower value.
5. Untick the box to go back to the untouched audio. With it off, the samples are not processed at all and cost no CPU.

What to expect

• The processing adds about 20 ms of delay (176 samples at 8 kHz, 959 at 48 kHz). The original is delayed by the same amount, so the blend stays aligned.
• Squelch and recordings use the cleaned audio, so a recording has the same 20 ms offset. See Audio recording gating.
• Steady hiss and noise between words drop strongly. Speech buried in noise stays hard to understand: no filter creates information that is not in the signal.
• It works at audio rates 8, 12, 16, 24 and 48 kHz; OnAir converts internally to 48 kHz in 10 ms frames.

Troubleshooting

• The box is missing: the demodulator is SSB, CW or a data mode. Switch to NFM, WFM or AM.
• Voices sound watery: lower the strength.
• Whistles or tones disappear: the filter treats non-speech as noise; turn it off for tones and data.
• No effect: check that the audio is not muted and that you listened in the mode that has the checkbox (Monitor audio does not use it).

Under the hood

The demodulated audio is brought to 48 kHz with a Kaiser low-pass, cut into frames of 480 samples, run through RNNoise and brought back. All memory is allocated when the receiver is made, so processing never allocates in the audio path. Strength changes are ramped across a block to avoid clicks.

Related

General receiver | Airband | Record audio and Monitor audio | Credits and contributors