Prosodic analysis of disfluent events in a corpus of university lectures

1 INESC-ID, 2 FLUL & 3 IST This paper describes our efforts towards the analysis of the prosodic properties (pitch, energy, and duration) of disfluencies, aiming both at a view of their global properties, and also at an analysis of their idiosyncratic behaviors. Underlying this task is the fact that disfluencies, e.g., filled pauses, prolongations, repetitions, substitutions, deletions, insertions, characterize spontaneous speech and play a major role in speech structuring [l]. For speech processing, the analysis of the regular patterns of those phenomena is crucial [2,3]. In automatic speech recognition (ASR), their identification accounts for more robust language and acoustic models [4] and even in text to speech synthesis (TTS), they are being modeled to improve the naturalness of synthetic speech [5]. Moreover, when combining ASR and TTS with machine translation systems, spontaneous speech translation still needs substantial improvements [6]. In our previous work for European Portuguese [7], we proposed that prosodic properties, mainly prosodic phrasing and contour shape, are essential to perform an evaluation task regarding fluency/disfluency distinctions. In this perspective, disfluencies may behave and even be rated as fluent communicative devices, when different segmental and suprasegmental aspects are monitored. We now aim at extending our study to a characterization of the prosodic parameters of disfluencies in a more quantifiable way. Specifically, our main goal is to verify if disfluent events have distinct prosodic properties. This work uses a subset of the LECTRA corpus [8], collected with the goal of transcribing university lectures for e-learning applications. The corpus has a total of 74h, of which 10h were multilayer annotated, including (besides other information) an orthographic tier, a morpho-syntactic tier, and a disfluency tier, annotated accordingly to [2]. This small subset corresponds to 5 speakers (the lecturers in each of the recorded courses). Table 1 shows the distribution of the different disfluencies for each of the speakers. al eti iou oop pmc Total