Feature Sets for the Automatic Detection of Prosodic Prominence

This work presents a series of experiments which explore the utility of various acoustic features in the classification of words as prosodically prominent or nonprominent. For this set of experiments, a 35,009 word subset of the Buckeye Speech Corpus was used [12]. This subset is divided across fifty-four segments of the Buckeye Speech Corpus. In a previous study, the words were transcribed for prosodic prominence by several teams, of sixteen naive native-speakers of English each, using the method Rapid Prosody Transcription developed in our prior work [10]. In the present study, we mapped the quasi-continuous valued prosody labels from the transcribed portion of the corpus to a binary prominence label. If at least one rater deemed a word prominent, it was labeled ‘prominent’ or otherwise it was labeled ‘nonprominent.’ 15,955 were labeled ‘prominent,’ yielding a baseline chance level of prominence assignment at 54.4%. 90% of the words were used in training the learning algorithms and the other 10% was used in testing. Several acoustic correlates are associated with prominence, including F0, duration, and intensity [1, 2, 4, 17, 3, 6, 15, 16, 11, 7, 14, 18]. The relative contribution that these play in speech recognition and in recognition by humans is well discussed in the literature [5, 18, 9, 13]. In the first set of experiments, Support Vector Machines (SVM) were used. SVMs were chosen because the task is a vector-input, class-label-output task, and SVMs do well at such tasks. Here, a set of 36 features was used, including both features known to be correlated to prominence and features not known to be correlated, such as the length of the pause after a word. The ten best-performing features were, in order, the minimum energy of the final vowel normalized by phones, the ratio of the energy of the current word to the next word, the post-word pause duration, the word duration normalized by phones, the maximum energy of the last vowel normalized by phone-class, the minimum value of f0 in the following word, the maximum energy of the stressed vowel normalized by phone-class, the stressed vowel duration normalized by phone-class, the minimum energy of the next word, and the maximum energy of the word. The classification accuracy was tested with related features clustered together into four groups: pause, duration, intensity, and pitch. The results are reported in table 1 For the second set of experiments, Hidden Markov Models (HMM) with three hidden