Simulating speech systems

Towards human-like spoken dialogue systems

“…This way, human-computer (wizard) and human-human data are collected from the same dialogue. A general prerequisite for Wizard-of-Oz studies is that users be under the impression that they are talking to a computer (Fraser & Gilbert, 1991, Wooffitt et al, 1997, as is the case in the studies above. The aim of studying human-like behaviour, however, may void this prerequisite.…”

“…Techniques to achieve this include instructing the wizard to misunderstand (Bernsen et al, 1998); using text filters that introduce insertion and deletion errors (Fraser & Gilbert, 1991); using a wizard to determine what utterances the system should even attempt to understand (Peissner et al, 2001); training wizards to speak without disfluencies and removing samples from the speech signal (Lathrop et al, 2004); providing the wizards with a limited selection of pre-recorded system prompts ; and limiting the wizard to making choices within a strictly defined dialogue model (Tenbrink & Hui, 2007). …”

A statistical approach to spoken dialog systems design and evaluation

“…The WOz technique (Fraser and Gilbert, 1991) allows the acquisition of a dialog corpus with real users without having a complete dialog system, for which the dialog corpus would be necessary. We chose human-wizard rather than human-human dialogs since people behave differently toward (what they perceive to be) machines and other people.…”

Statistical framework for a Spanish spoken dialogue corpus

“…This acquisition was only restricted at the semantic level (i.e., the ac-quired dialogues are related to a specific task domain) and was not restricted at the lexical and syntactical level (spontaneous-speech). In this acquisition process, the semantic control was provided by the definition of scenarios that the user had to accomplish and by the WoZ strategy (Fraser and Gilbert (1991)), which defines the behaviour of the acquisition system.…”

“…The acquisition of the Dihana corpus was carried out by means of an initial prototype, using the Wizard of Oz (WoZ) technique (Fraser and Gilbert (1991)). This acquisition was only restricted at the semantic level (i.e., the ac-quired dialogues are related to a specific task domain) and was not restricted at the lexical and syntactical level (spontaneous-speech).…”