Synthesized lengthening of function words - The fuzzy boundary between fluency and disfluency

As [1]’s model of speech production suggests, speakers sense upcoming difficulties and can correct them before uttering. A reasonable strategy to bridge resulting gaps is to prolong the words in the articulatory buffer [2]. This often buys enough time to correct the issue, resulting in standalone disfluent lengthening, after which fluency is resumed [3]. In case of more severe difficulties, the lengthening may be followed by other disfluencies such as silent or filled pauses or repetitions. Similar hesitation strategies might be useful in automatic speech production, e.g. for spoken dialogue systems that interact with human users and typically face a variety of challenges in natural language understanding and generation. Lengthening is an ambivalent phenomenon in speech that seems to be located at the fuzzy boundary between fluency and disfluency. It regularly occurs before phrase boundaries [4][5] and besides constitutes a common hesitation disfluency. Some disfluencies consist of lengthening [3] only, and some lengthenings appear so subtle that they pass unnoticed [6][7]. We assume that these characteristics of lengthening make it a key component in spoken dialogue systems that are capable of producing disfluencies, as they enable to buy a variable amount of time whilst being unobtrusive to the listener [6]. It is not yet known, however, how much synthetic lengthening is acceptable and how lengthening influences the user’s interaction with the system. To address these issues, this study tests the effects of step-wise increases of synthesized lengthening on user ratings and interaction speed.