mvdr_2mic

wakeword driven MVDR beamformer

# 1. Install dependencies
sudo apt-get update
sudo apt-get install -y build-essential libasound2-dev git python3

# 2. Enable the virtual audio loopback
sudo modprobe snd-aloop
echo "snd-aloop" | sudo tee -a /etc/modules

# 3. Clone and build
git clone https://github.com/rolyantrauts/mvdr_2mic.git
mkdir mvdr_2mic/build
cd mvdr_2mic/build
cmake ..
make -j 3

# 4. Run the engine (Wait for Python IPC commands)
./mvdr_engine --mic hw:1,0 --out plughw:Loopback,0,0 --debug
# Open another terminal
arecord -D hw:Loopback,1,0 -f S16_LE -r 16000 -c 1 test_beam.wav

:~/mvdr_2mic/build $ ./mvdr_engine --help
=== MVDR PFFFT Standalone Engine (IPC Controlled) ===
Usage: ./mvdr_engine [OPTIONS]

Options:
  -h, --help           Show this help message and exit.
  --mic <hw:X,Y>       Set ALSA capture device (Default: hw:1,0).
  --out <hw:X,Y>       Set ALSA playback device for processed audio (e.g., plughw:Loopback,0,0).
  --in <file.wav>      Bypass ALSA mic and process a raw 16kHz 2-channel WAV file directly.
  --mic-spacing <m>    Distance between microphones in meters (Default: 0.058).
                       [!] MIN: ~0.03m (30mm) to overcome hardware phase variance (WNG limit).
                       [!] MAX: ~0.042m (42mm) for pure voice band without spatial aliasing.
                       (Note: The default 0.058m leverages an internal 2.9kHz anti-aliasing guard).
  --doa-history <n>    Number of 16ms frames to keep for the LOCK average (Default: 100 = 1.6s).
  --measure-clock      Run a 10-second hardware clock drift test and exit.
  --debug              Enable live terminal metering, CSV telemetry, and raw WAV dumping.

IPC Control:
  This engine listens on UNIX socket: /tmp/mvdr_ipc.sock
  Commands:
    'VAD 1' : Start steering beam (Voice detected)
    'VAD 0' : Stop steering, learn background noise
    'LOCK'  : Analyze history buffer, lock beam direction
    'RESET' : Unlock beam, return to background noise mode
    'SET X' : Manually force beam to X degrees (e.g., 'SET 45')

python3 mvdr_test.py
=== MVDR Manual Pilot ===
Act as the neural network. Send commands to the C++ engine.
Commands:
  1 - Send 'VAD 1'  (Starts DOA Steering)
  0 - Send 'VAD 0'  (Stops Steering, learns background noise)
  L - Send 'LOCK'   (Evaluates the 1.6s history buffer and locks)
  R - Send 'RESET'  (Unlocks and returns to AUTO)
  Q - Quit
-------------------------
Pilot Command [1, 0, L, R, Q]: 

Use a wakword for VAD.

The models from https://github.com/rolyantrauts/BoWWClient/tree/main/models
https://github.com/rolyantrauts/BoWWServer_x86/tree/main/models Trained in https://github.com/rolyantrauts/wakeword_dataset/tree/main/gkws
or https://github.com/rolyantrauts/bcresnet have a 'noise' class that IMO is a better term than 'silence' Its an inverse VAD but that is really what you want to know that this is not voice. Its much less compute from a model you are already running, than external vad.