15th International Congress of Phonetic Sciences (ICPhS-15)

Barcelona, Spain
August 3-9, 2003

Segmenting the Speech Chain into Tone Units: Human Behavior vs. Automatic Processes

Francesco Cutugno, Leandro D'Anna

Università degli studi di Napoli "Federico II", Italy

A tool aiming at the automatic segmentation of connected speech into tone units has been developed. The program analyses various acoustic parameters (pauses, voicing breaks and restarts, energy patterns) in order to extract prosodic coherent units. As usual in these cases, in order to evaluate the reliability of the system, a comparison between the software outputs and data deriving from the manual analysis performed by one or more human experts has been considered. This paper will describe a procedure used to determine the accordance degree for the segmentation in tone units of connected Italian. A group of 11 expert Italian prosodists were asked to work on a common set of utterances derived by a map-task corpus, the resulting accordance degree is, on average, less than 75%, indicating a relatively high level of division among the experts about tone unit definition. Their decision were compared with the automatically produced ones. Accurate observation were conducted on the acoustic features available on the unit boundaries both in the case of maximum accordance and in that of minimum accordance. Outcomes of this analysis have been used to improve the algorithms' reliability.

Full Paper

Bibliographic reference.  Cutugno, Francesco / D'Anna, Leandro (2003): "Segmenting the speech chain into tone units: human behavior vs. automatic processes", In ICPhS-15, 1233-1236.