Herme the Collection of Natural Speech Data

This paper describes our approach to the collection of ‘natural’ (i.e., representative) data from spoken interactions in a social setting in the context of the development (through time) of expressive speech synthesis. Over the past ten years or so, we have collected several corpora of unprompted social conversations that illustrate the ‘contact’ element of speech that was lacking in many of the corpora collected by use of a specific ‘task’ with paid participants. The paper discusses the technical and ethical issues of collecting such spoken material, and highlights some of the problems we have encountered in the processing of this much-needed data. Through the use of attractive conversational devices, we have found that natural human curiosity, and an element of social programming combine to provide us with a rich source of material that complements the task-based collections from paid informants.