 赵昌
|
e589cdd874
|
feat(acoustic): Recover full acoustic analyzer module with buildable demo
- Recover all 50+ source/header/script files from conversation history
- Regenerate synthetic binary dataset (ambient vs threat)
- Retrain CNN-GRU classifier: 15 epochs, 90.4% train / 100% val accuracy
- Export ONNX model with MaxPool1d workaround for opset 13 compatibility
- Fix recovered file API mismatches (AudioBuffer, Pipeline, demo_offline)
- Replace kiss_fft with inline DFT in fft_utils.cpp (no external lib needed)
- Add missing #include <memory> to PIMPL headers
- Fix ONNX input layout: [batch, time_frames, n_mels]
- Downgrade ORT_API_VERSION 20 -> 17 for local DLL compatibility
- Simplify build_demo.bat for Windows MinGW (no yaml-cpp dependency)
- Verify demo_offline.exe on validation set: 20/20 correct classifications
|
1 day ago |
p82feo7wg