Corpus Development Activities at the Center for Spoken Language Understanding

Corpus Development Activities at the Center for Spoken Language Understanding Ron Cole, Mike Noel, Daniel C. Burnett, Mark Fanty, Terri Lander, Beatrice Oshika, Stephen Sutton Center for Spoken Language Understanding Oregon Graduate Institute of Science and Technology Portland, Oregon 97291 ABSTRACT This paper describes eight telephone-speech corpora at vari- ous stages of development at the Center for Spoken Language Understanding. For each corpus, we describe data collection procedures, methods of soliciting callers, protocol used to col- lect the data, transcriptions that accompany the speech data, and the expected release date. The corpora are available at no charge to academic institutions. 1. INTRODUCTION The Center for Spoken Language Understanding (CSLU) collects and transcribes telephone-speech data to enable re- search activities at CSLU and elsewhere. Corpus develop- ment a~ctivities are performed by four full-time staff, aided by graduate students and paxt-time employees. In 1994, we an- ticipate collecting and transcribing speech from 10,000 callers in twenty languages. Corpus development activities are sup- ported by industrial memberships and research grants. Corpus development activities at CSLU include: (a) collect- ing telephone speech data in different languages; (b) tran- scribing speech at word and phonetic levels; (c) developing and documenting transcription conventions for each level; (d) measuring the level of agreement among transcribers; (e) de- veloping interactive speech tools for labeling; (f) distributing the speech corpora to academic institutions free of charge; and (g) placing speech tools and labeling conventions in the public domain for use by others. In this section, we present some general information about our corpus development activities. In the following sections, we will describe individual corpora. Data Collection. telephone-speech data are collected over analog and digital telephone lines. Prior to November, 1993, speech data were collected over analog lines using several Gradient Technology Desldabs. Since November, 1993, the majority of our data has been collected using a 24 channel T1 line connected to three LINKON FC3000 Communica- tion Boards. We are also using an Apple GeoPort Telecom Adapter connected to a Macintosh Quadra A]V to collect analog speech data for one of the corpora to be described. aligned phonetic level. Non-time-aligned word level tran- scription involves producing an orthographic representation of the utterance, including indications of extra-speech events such as breathes or lip smacks, without time markings. Time- aligned word level transcription provides the same ortho- graphic transcription augmented with time alignment mark- ings. Time-aligned phonetic transcription involves aligning phonetic symbols to the acoustic signal. A precise description of the conventions used for all levels of labeling, including a complete list of all phonetic labels for each language, is presented in the CSLU conventions document[l] Transcription Reliability. We are conducting experiments to determine the level of agreement among labelers. In these experiments, CSLU staff and professional phoneticians axe using Worldbet [5] to transcribe the same intervals of speech. Initial results for English indicate overall agreement of ap- proximately 80% across all labels, ranging from approxi- mately 70% for vowels to greater than 90% for stops and nasals. Speech Tools. The OGI …

Content truncated to 500 words