Spontaneous Speech Collection for the CSR Corpus

SPONTANEOUS SPEECH COLLECTION FOR THE CSR CORPUS Jared Bernstein and Denise Danielson SRI International Menlo Park, California 94025 1. ABSTRACT As part of a pilot data collection for DARPA's Continuous Speech Recognition (CSR) speech corpus, SRI International experi- mented with the collection of spontaneous speeoh material. The bulk of the CSR pilot data was read versions of news articles from the Wall Street Journal (WSJ), and the spontaneous sentences were to be similar material, but spontaneously dictated. In the first pilot portion of the data collection, twelve subjects including nine journalists were located, and instructed in how to dictate using the data collection hardware and software at SRI. These talkers pro- duced 1280 spontaneous sentences. In general, compared to read material, the spontaneous material took about two to three times more subject time to produce and about four times more experi- menter time to produce, package, and ship. The paper provides details on the materials, subjects and procedures used in the study, and it describes the results in terms of speaker reaction and data production. The methods described are sufficient to collect fluent spontaneous recordings at a predictable rate. The spontaneous material differs in several characteristics from WSJ material; para- graphs and sentences tend to be longer, more word types are used, and by most measures, the material is more variable. 2. INTRODUCTION The CSR (Continuous Speech Recognition) Corpus collec- tion can be considered the successor to the Resource Man- agement (RM) corpus[l], it focuses on the further development of speech recognition technology toward larger or open vocabularies, speaker and task indepen- dence, and is moving toward spontaneous speech. The default task in the pilot collection has been dictation of newspaper articles as if for the Wall Street Journal (WSJ). Thus, the largest part of the effort to collect a pilot version of the CSR corpus has been recording people reading selected short passages from the WSJ itself. The pilot CSR corpus was designed, however, such that a significant por- tion of the material was to be spontaneous and a subset of the speakers who read WSJ texts also were asked to dictate spontaneous articles in the WSJ style. This paper describes the methods used in the collection of the spontaneous portion of the CSR corpus. 3. METHOD User Interface. The speech data collection was performed with user interface software designed by Mike Phillips for collection of read speech [2]. MIT provided this software to SRI, where it was slightly modified for use in collecting spontaneous speech. The interface requires a talk button to be pushed and held down in order to record speech; another button must be explicitly pushed to accept the sentence. The most recent sentence is always available for playback. Material Selection. Several issues seemed important in the selection of materials to be used in story generation. 1. It seemed appropriate to select material that would match the content of the WSJ, its vocabulary and top- ics. 2. To ensure that speech is tngy spontaneous and not …

Content truncated to 500 words