Integrating prosodic information into recurrent neural network language model for speech recognition

Prosody is a kind of cues that are critical to human speech perception and comprehension, so it is plausible to integrate prosodic information into machine speech recognition. However, as a result of the supra-segmental nature, it is hard to integrate prosodic information with conventional acoustic features. Recently, RNNLMs have shown to be the state-of-the-art language model in many tasks. We thus attempt to integrate prosodic information into RNNLMs for improving speech recognition performance based on rescoring strategy. Firstly, three word-level prosodic features are extracted from speech and then passed to RNNLMs separately. Therefore RNNLMs predict the next word based on prosodic features and word history. Experiments conducted on LibriSpeech Corpus show that the word error rate decreases from 8.07% to 7.96%. Secondly, prosodic information is combined on feature-level and model-level for further improvements and word error rate decreases 4.71% relatively.