PROSODIC PROMINENCE DETECTION IN SPEECH Fabio Tamburini CILTA/DEIS - University of Bologna - Italy f.tamburini@cilta.unibo.it ABSTRACT This paper presents work in progress on the automatic detection of prosodic prominence in continuous speech. Prosodic prominence involves two different phonetic features: pitch accents, connected with fundamental frequency (F0) movements and syllable overall energy, and stress, which exhibits a strong correlation with syllable nuclei duration and high-frequency emphasis. By measuring these acoustic parameters it is possible to build an automatic system capable of correctly identifying prominent syllables with an agreement with human-tagged data comparable with the inter-human agreement reported in the literature. These results were achieved without using any information apart from acoustic parameters. 1 INTRODUCTION The study of prosodic phenomena in speech is a central topic in language investigation. Speakers tend to focus the listener's attention on the most important parts of the message, marking them by means of such phenomena. As outlined in Beckman & Venditti [4], a precise identification of such phenomena helps to disambiguate the meaning of some utterances. It is also a fundamental step for the automatic recognition of spontaneous speech, and enhances the fluency and adequacy of automatic speech-generation systems. Moreover the construction of large annotated language resources, such as prosodically tagged speech corpora, is of increasing interest both for research purposes and for language teaching. One of the most important prosodic features is prominence: a word or part of a word made prominent is perceived as standing out from its environment [23]. A better understanding of how prominence is physically accomplished is a basic step in the construction of tools capable of automatically identifying such phenomena. This paper presents work in progress on the construction of a system for the automatic detection of prosodic prominence features in speech using only acoustic/phonetic parameters and cues. Following Beckman's [3] phonological view, further developed by Bagshaw [1, 2], syllables that are perceived as prominent either contain a pitch accent or are somehow "stressed". On the acoustic/phonetic side, the accomplishment of such features has to be strictly correlated with acoustic parameters. As well as the works already cited, there are many studies [15, 16, 17], suggesting that some of the main acoustic correlates of prominence are pitch movements (strictly connected with fundamental frequency - F0), overall syllable energy, syllable duration and spectral emphasis. The work presented here is divided into two separate steps: the first step involves the automatic identification of syllable-nuclei boundaries to reliably measure the duration feature, while the second one concerns the identification of prominent syllables by means of acoustic measurements. This paper will report on the first experiments conducted on the whole system. The data set used in these experiments is a subset of the DARPA/TIMIT acoustic-phonetic continuous speech corpus, consisting of thousands of transcribed, phone- segmented and aligned sentences of American English. In this study the TIMIT annotations are used only for measuring the system performances, not for prominence detection. Several studies have been conducted in this field for building automatic systems capable of reliably identifying either one acoustic correlate of prominence [5, 7] or a complete set of prosodic parameters [2, 6, 24]. These latter studies, involved in the construction of a complete prosody identification system, rely on additional phonetic information such as phone labelling and/or utterance transcriptions. Despite the quantity and quality of studies on this topic, it seems that the automatic and reliable detection of prosodic prominence, without providing phonetic information, is still an open question. 2 THE ACOUSTIC PARAMETERS In the following subsections, each acoustic parameter involved in this study is considered. All acoustic parameters must be normalised to some extent to avoid the natural variations among different speakers. The specific normalisation procedures applied to each parameter will be described. 2.1 Duration The linguistic theories of prosodic prominence listed above tend to consider syllable duration as one of the fundamental acoustic parameters for detecting syllable stress. Unfortunately the automatic segmentation of the utterance into syllables is a complex task; in [9] we can find a survey of syllable segmentation algorithms. None of these methods seem to perform well when applied to continuous speech. For these reasons, an alternative duration measure for prosodic prominence detection should be introduced. One possible measure seems to be the duration of syllable nucleus. Considering some utterances taken from the TIMIT corpus and comparing the duration of the syllable nucleus with the duration of the entire syllable, with respect to prominence, and approximating the 0-7803-7946-2/03/$17.00 ©2003 IEEE. ISSPA 2003