The results: about 90% accuracy has been achieved, so it is satisfying.
- CNN on musical instruments recognition
- CNN on human activity in my home
- CNN on key word detection
- DNN on key word detection
- CNN on my life environment
The trained CNN model will work OK if the following condition is satisfied:
- use the same MEMS mic with the same filter setting (incl. the same frequency response) on the edge device (STM32L4)
- same environment
- less surrounding noise
- same window setting