The effect of pitch, intensity and pause duration in punctuation detection

The purpose of this research is to automatically detect punctuation in speech using only prosodic cues. We aim to integrate prosodic elements such as pauses, changes in f₀ and amplitude range, into an Automatic Speech Recognition engine in order to generate punctuation for read speech, without taking the context of the sentences into consideration. We trained acoustic models of the prosodic features of two Punctuation Marks (PMs): full-stop and comma, which we assume have distinct prosodic characteristics. A Neural Network was used to estimate the weights assigned to each prosodic feature that corresponds to a particular PM, later to be used by a PM classifier. Results show that 87% of full-stops were detected, with only 14% false alarms. Nevertheless, since most commas are realized with no pitch breaks, only 54% of the commas were detected, with 35% false alarms. Our results support the hypothesis that acoustic-prosodic cues provide useful evidence about phrases.