An automatic transcription system of hearings in Italian courtrooms

Abstract

This paper describes and discusses the recognition results obtained using the automatic transcription system developed in our labs, after having adapted it to the judicial domain. The performance has been evaluated on field audio data, formed by about 7 hours of multiply tracks recordings, acquired in two different dates in the Court of Naples. Different sets of acoustic and language models have been used and compared in the system, providing results (word error rate is around 40%) that are in line with those obtained on other comparable Automatic Speech Recognition (ASR) tasks (e.g. meeting transcription)and that leave room for future investigations