Creating a speech corpus with semi-spontaneous , parallel conversational and clear speech Tech Report : CSLU-11-003

Our goal is to collect a speech corpus for the purpose of studying intelligibility and acoustic differences between the conversational and clear speech styles. The ideal corpus has the following properties: (1) speech has been produced spontaneously as part of a communicative interaction, as opposed to having been read to an imagined interlocutor; (2) entire identical utterances, or large parts of utterances, are available in both conversational and clear speaking styles, also known as parallel recordings; and (3) utterances comprehensively and systematically cover the space of prosodic and phonetic features. We call the spontaneous (i. e. non-read) elicitation of speech with highly anticipated content (established through a given task) semi-spontaneous. We now discuss these desirable properties in more detail.