I had the same condition: High acc and low vad_acc.
It was because the parameter of Keras.model.fit, validation_split.
This will separate the last section of data as validation data. Therefore, if your data was in order, your validity data will be in the same case. Try to shuffle the training data.
