Hear the AI remove battlefield noise
The neural network runs on this device: no server, nothing uploaded, and it keeps working offline once loaded.
- 1Put on earphones or headphones
- 2Press Start and allow the microphone
- 3Speak, add noise, and switch between A and B
measuring…
What you hear
Tap to switch while you speak. Both paths are time-aligned, so only the noise changes.
Add battlefield noise
Mixed into your microphone signal before the AI sees it.
Speak for a few seconds: the noise level is matched to your voice.
Lower = louder noise. The level applies to all active noise together.
Start the demo first
The radio controls work on your live microphone.
When should the radio transmit?
A conventional voice switch keys on loudness, so gunfire keys it. The AI switch keys only on speech.
Try it: fire the gunshot pad without speaking. The conventional switch lights up on the gunfire; the AI one stays dark until you talk.
Start the demo first
Speed and latency are measured on this device while it runs.
Measured on this device
Live numbers from the model running here.
Model
GTCRN fine-tuned on defence noise, 23.7 K trainable parameters.
The same network stored two ways. INT8 covers the inner convolution and linear layers; the input and output layers, GRUs and transposed convolutions stay FP32, because quantising them cost 2 dB. At this model size INT8 is neither smaller nor faster. It is here to show quantised inference running on a phone.