A recent trend has been to use a language model to do the scoring: https://arxiv.org/abs/1904.09675