Using various types of multimedia resources to train system for automatic transcription of Czech historical oral archives

Abstract

Historical spoken documents represent a unique segment of national cultural heritage. In order to disclose the large Czech Radio audio archive to research community and to public, we have been developing a system whose aim is to transcribe automatically the archive files, index them and make them searchable. The transcription of contemporary (1 or 2 decades old) documents is based on the lexicon and statistical language model (LM) built from a large amount of recent texts available in electronic form. From the older periods (before 1990), however, digital texts do not exist. Therefore, we needed a) to find resources that represent language of those times, b) to convert them from their original form to text, c) to utilize this text for creating epoch specific lexicons and LMs, and eventually, d) to apply them in the developed speech recognition system. In our case, the main resources included: scanned historical newspapers, shorthand notes from the national parliament and subtitles from retro TV programs. When converted into text, they allowed us to built a more appropriate lexicon and to produce a preliminary version of the transcriptions. These were reused for unsupervised retraining of the final LM. In this way, we significantly improved the accuracy of the automatically transcribed radio news broadcast in 1969-1989 era, from initial 83 % to 88 %. © 2013 Springer-Verlag.