Improve Precision of Negative class in Neural Network Output

Question

I have trained CNN for predicting the audio tone (Negative, Neutral, Positive) and i have 4 different CNN models with slightly different architecture and datasets. These models