NGN — Voice Engine for VOICEN Linear 4-Mic Array Kit

English | 简体中文

Real-time voice front-end for the VOICEN Linear 4 Mic Array Kit (AC108 4-mic uniform linear array, 40 mm spacing, on a NanoPi Air). It runs as a daemon on the board and provides:

• AEC — echo cancellation (21.9 dB synthetic / 10–13 dB real-music ERLE)
• DOA — direction finding (±3° online on real sources)
• Beamforming — DSB output, 6.2 dB array gain
• Virtual sound cards — any ALSA app can record the processed mic (arecord -D voice_mic_shm) and play back with automatic echo reference (aplay -D voice_spk_shm)

The full pipeline runs in real time on a single core (rtf ≈ 0.4–0.5).

Quick start

On the board (or any Linux host for the algorithm tests):

make -C algo && (cd algo && ./test_algo)      # algorithm unit tests
make -C engine && ./engine/test_engine.sh     # engine suite
make -C plugin && sudo make -C plugin install # ALSA plugins
test/test_shm.sh                              # IPC integration suite

Run the engine:

engine/voice_engine --ipc shm --mic-order 0,3,2,1 --stats-interval 10

Then from any application:

arecord -D voice_mic_shm -f S16_LE -r 16000 -d 10 rec.wav   # record (echo-free)
aplay  -D voice_spk_shm music.wav                          # play (echo-cancelling)

Deploy from a dev machine: algo/sync_build.sh and engine/deploy.sh (target defaults to ssh host nanopi, path /home/i/ngn).

Documentation

• docs/PIPELINE.md — architecture, C ABI, internals (for developers)

License

MIT