At the training phase，I Collected 50 wave files，then feed into the voice model，after train on web，Accuracy 93.85% was got. Just "On/OFF/noise” voice is needed. Then I deploy a c++ library outside，integrate it into a PSoC6（MCU，not AI chip） board of Infineon，then test. At the application phase, I find the classifier's error rate is high, not like the 93.85% resulted on the website. What's the reason? 1、How many samples are needed to train it? Ten thousand？ 2、Any Parameters can be adjusted to optimize it? 3、What cause the accuracy changed from web to tflite library?