Phone Elasticity in Disfluent Contexts

Disfluencies in speech are instances of hesitation or correction that affect both the speaker and the listener. Typical surface forms of disfluencies are filled pauses (such as uh, uhm), silences at places in the utterance where syntax would not predict them, or repetitions of parts of the utterance (like (I mean + I mean) we should go). Disfluencies carry a folk notion of erroneousness or badness. Voice coaches often advise to eliminate disfluencies such as filled pauses from speech. Research in the past decades began to put disfluencies into a more positive light, as it was discovered that they provide the listener with important meta-information about the current speech from their interlocutor. Listeners can infer difficulties of reference or problems of forming the right speech plan when they hear the speaker hesitate (See [1] for an overview). We largely follow the view of [2], who formulated the new perspective on disfluencies as them being solutions, not problems. Recent studies started investigating the question if disfluencies are suitable for human-machine communication as well, cf. [3]. One of the possible places of applicability is incremental spoken dialogue systems. Compared to non-incremental systems, these systems can prepare their response to human speech input in real-time while the user is speaking and can thus respond with only milliseconds delay, cf. figure 1. This, however, can cause problems such as running out of things to say or uttering erroneous output which needs correction. It is an open question whether disfluencies can remedy these issues by equipping incremental spoken dialogue systems with human features of dialogue management in order to solve these problems in addition to sounding more natural. With the aim in mind to improve the quality of incremen-