BEST RESULT · EVALUATION SET
SNR
20.2dB
target > 15TARGET
STOI
0.88
target > 0.85TARGET
PESQ
2.60
target > 2.5TARGET
Latency
20ms
wireless linkTARGET
These results are measured on moderate noisy conditions. In sever noise, the SNR gain is > 18 dB, STOI > 0.8, PESQ > 2.3 and this is going to improve with increase in training data.
01The problem
Three things make this hard.
- 01
Noise doesn't stand still.
Gunfire, rotors and sirens change faster than classical filters can track.
- 02
Filters distort speech.
Spectral subtraction and Wiener filtering smear consonants and add artefacts when they over-subtract.
- 03
The hardware is tiny.
It all has to run in real time on a microcontroller, within a 10 ms frame.
02Signal chain
How it works
Every 10 ms frame walks the same chain. A gate decides whether it needs the neural stage at all.
- Prefilter
- The reference mic hears the world, the primary hears the mouth. An NLMS filter subtracts one from the other and steady noise falls away.
- Gate
- Above +5 dB the DSP has already done the job, so the frame skips the network and saves the cycles.
- Network
- Below that, a quantisation-aware GRU sets a gain per band every 10 ms and clamps what the filter couldn't follow.
03Hardware
The prototype
A sealed, printed enclosure carrying the MCU, the reference mic and the battery. Drag to look around it.
DRAG TO ROTATE
Built to be heard in the loudest places.