Spontaneous Speech Collection for the ATIS Domain with an Aural User Feedback Paradigm

Spontaneous Speech Collection for the ATIS Domain with an Aural User Feedback Paradigm Christine Pao and Jay Wilpon AT&T Bell Laboratories 600 Mountain Ave. Office: 2D-464 Murray Hill, NJ 07974 ABSTRACT This paper describes the AT&T ATIS data collection system, with emphasis on the development of the speech-in, speech- out interaction paradigm. The ATIS task involves providing air travel information to a user in the context of a interactive dialogue. Under the AT&T interaction paradigm, informa- tion retrieved from a travel information database in tabular form is automatically transformed into sentences, which are spoken to a user by a speech synthesizer. To date, we have collected over 1800 sentences from subjects who used the sys- tem to solve travel planning scenarios. We present a compar- ison of the ATIS data collected at AT&T with the ATIS data collected at other sites (BBN, CMU, MIT, and SRI), and discuss what we have learned in this prehminary effort. 1. INTRODUCTION In support of our research toward developing telephone- based spoken language systems, we have joined BBN, CMU, MIT, and SRI in collecting speech and language data for the ATIS (Air Travel Information Service) do- main. The task of a spoken language system can be broken into three parts. First, the system must under- stand and interpret what the user says in the context of the human-machine dialogue. If the user's request is understood, the system must retrieve the requested in- formation. The information source for the ATIS domain is a relational database that represents a 10-city subset of the OAG, or Official Airline Guide. Finally, the sys- tem must convey the retrieved information to the user in an appropriate format, or provide some other coop- erative response, such as a request for more information or an error message. In the AT&T ATIS data collec- tion system, the first and second portions of the task are handled by components of the MIT ATIS system [1]. Our efforts have been directed toward dealing with the third part of the task: information presentation and sys- tem feedback under a speech-in, speech-out interaction paradigm. Our goal is to begin to address the problems of dialogue and information control which will affect the performance of interactive spoken language systems. In this paper, we will first describe our data collection system and data collection procedure. Then, we will present analyses of AT&T ATIS data and of ATIS data collected at other sites. Finally, we will discuss what we have learned in this preliminary effort. 2. DATA COLLECTION As at most of the other sites collecting ATIS data, data were collected from subjects' interactions with a partially simulated spoken language system. As in the MIT data collection setup, a human experimenter is substituted for the speech recognition component of the system to provide a transcription of the subject's speech for the natural language (NL) component. In this section, we discuss the system's development, describe the system hardware, and describe the collection procedure. 2.1. System Development The AT&T data collection …

Content truncated to 500 words