RADIOLOGICAL REPORTING BY SPEECH RECOGNITION: THE A.Re.S. SYSTEM B. Angelini, G. Antoniol, F. Brugnara, M. Cettolo, M. Federico, R. Fiutem and G. Lazzari IRST-Istituto per la Ricerca Scientifica e Tecnologica I-38050 Povo (Trento), Italy ABSTRACT Radiological reporting has already been identified as a field in which voice technologies can prove to be very useful. Recent progress in automatic speech recogni- tion and in hardware and software technology makes it possible to build large-vocabulary, continuous speech, speaker-independent, real-time systems. In this paper a dictation system for radiology re- porting, the A.Re.S. system, is presented. A.Re.S. is a “software only” system which runs in real-time on an HP 715 workstation. It relies on an asynchronous and multi-process architecture in which speech decoding is performed by processes in pipeline. System requirements and architecture will be de- scribed, together with the results of a preliminary eval- uation based on three months of on-site testing. I. INTRODUCTION Recent progress in Automatic Speech Recognition (ASR) and in hardware and software technology makes it possible to build large-vocabulary, real-time, speaker-independent systems. Medical document gen- eration presents features which render it a good ap- plication field for ASR [5, 6], since the employed lan- guage is typically standardized and the vocabulary size is manageable. In the particular field of radiological reporting, for example, computer dictation allows a ra- diologist to dictate naturally a report while looking at the X-ray photograph. In this paper an application of speaker independent, continuous speech, and real-time ASR for radiology is presented. The system, named A.Re.S. (Automatic Reporting by Speech), covers a vocabulary of 6,000 words, related to the emergency examinations domain. A prototype of A.Re.S. has been installed at S. Chiara Hospital 1 and is daily used by physicians. 1 S. Chiara Hospital is a regional structure that serves a large community of about 100,000 people; its Radiological Department performs more than 150,000 examinations a year. II. THE A.Re.S. PROJECT 2.1 Radiological Reporting A radiological report is the typed document that de- scribes and summarizes the physician’s observation de- ducted from an X-ray photograph. In the daily routine, reports are usually recorded by physicians on tape cassettes and then transcribed by secretaries. In emergency medicine, where report gen- eration time is critical, a secretary instead types the report under physician’s dictation. In both radiology and emergency medicine, report generation can be automated using an ASR system, which supplies a draft text transcription that needs only to be revised by the secretary or by the physician her/himself. 2.2 Project Overview The A.Re.S. project has been developed at IRST in collaboration with the Radiological Department of S.Chiara Hospital, Trento. It started in 1990 with a feasibility study to identify requirements and propose technologies [3]. The first prototype [4], delivered in May ’92, allowed the dictation of reports related to chest examinations (2,500 words of vocabulary), with a small pause between words, and in speaker depen- dent modality. It ran almost in real-time on a HP 720 workstation under Unix, and used a DSP board to perform feature extraction. The current A.Re.S. pro- totype overcomes the limitation of isolated word dicta- tion by employing a speaker independent and continu- ous speech recognition engine. 2.3 System Description Following the distinction between routine and emer- gency reporting, A.Re.S. provides two operation modalities: batch and interactive. In batch modality, A.Re.S. follows a tape recorder metaphor. A digital voice recorder is simulated which resembles a normal tape recorder with the exception that it does not record pauses between words or sen- tences. The report, dictated through a microphone, is stored in a logical cassette and is transferred in a waiting queue of cassettes, which represents the set of recordings confirmed by the user and to be processed