too much false positives in wake word detection

Question

I tried to implement a wake word detection using TensorFlow. But it detects most of the noises as negatives.

The negative word dataset size is 679, taken out from [ht