Improved Transcription of Czech Parliament Speeches by Acoustic and Language Model Adaptation

The aim of this work is to improve the accuracy of our spoken broadcast transcription system in the task of Czech parliament speeches recognition. To achieve this goal, we propose several approaches for adaptation of both acoustic and language models of our system: a new two step unsupervised speaker adaptation strategy is presented to improve the former model while the latter one is created from a text corpus mixed properly from both general (2.6 GB of Czech newspaper texts) and domain specific data (181 MB of parliament speeches). Our experimental results show that the combination of both adaptation approaches leads to near 30% relative reduction of WER in comparison with the baseline speaker independent (SI) system operating with a general language model.