Chapter 7 Speech Recognizer, HMM Score very low

[dowenrei]

I understand that The MFCC of each category of fruit is used to built a HMM model respectively.
This line of code
score = hmm_model.get_score(mfcc_features) are used to score the model,
however the score turns out to be large and negative even though I used to predict the training audio , why is it so?
For example, the apple HMM predicts that the score of apple01.wav is -4183.