Fully automated approach to broadcast news transcription in Czech language

In the paper(1) we propose a complete scheme for automatic transcription of Czech TV news. The scheme first removes the music and noisy parts, then makes segmentation of the speech signal into speaker turns and consequently tries to decode and transcribe single utterances. We employ our own recognizer recently operating with a 200K-word lexicon and with a bigram language model. The overall recognition rate achieved on all the test data was 71.53%, that obtained on the read parts was 82.72%. The most serious recognition errors occur mainly in the segments that contain background music or extremely loud noise.