Modeling Word Duration for Better Speech Recognition

We describe a new method of modeling duration at word level. These duration models are easily trained from the acoustic training data and can be used to rescore N−best lists of recognition hypotheses. The models capture some of the well known durational effects such as prepausal lengthening. They incorporate a simple back off mechanism to handle unseen words during rescoring. Experiments with various large vocabulary conversational speech recognition (LVCSR) evaluation sets showed consistent improvements of 0.7−1.0% in word error rate (WER).