A combined punctuation generation and speech recognition system and its performance enhancement using prosody

摘要

A punctuation generation system which combines prosodic information with acoustic and language model information is presented. Experiments have been conducted for both the reference text transcriptions and speech recogniser outputs. For the reference transcription, prosodic information of acoustic data is shown to be more useful than language model information. Several straightforward modifications of a conventional speech recogniser allow the system to produce punctuation and speech recognition hypotheses simultaneously. The multiple hypotheses produced by the automatic speech recogniser are then re-scored using prosodic information. When the prosodic information is incorporated, the F-measure (defined as harmonic mean of recall and precision) can be improved. This speech recognition system including punctuation gives a small reduction in word error rate on the 1-best speech recognition output including punctuation. An alternative approach for generating punctuation from the un-punctuated 1-best speech recognition output is also proposed. The results from these two alternative schemes are compared.