How to get sentence probability from CTCLoss?

I need to compare probabilities of two sentences in an ASR. I have the logprobability matrix from the accoustic model and I want to use the CTCLoss to calcuate the probabilities of both sentences.
Does the CTCLoss return the negative log probability of the sentence? Or does it return pure probability of the given sentence? It would make sence to me that it would return negative log probabilities. However, I am receivin relatively small values (less than 1) for long sentences. I would expect that long senences would have very low log-likelihood and therefore very large negative log-likelihood (like thousands or so).