An investigation into VTLN for improved transcription of Czech broadcast programs

This paper deals with the Vocal Tract Length Normalization (VTLN) method. The aim is to investigate the best way how to utilize this technique for improving recognition accuracy of a LVCRS system that has been developed for broadcast program transcription at our lab in recent years. For this purpose, VTLN is evaluated experimentally in several configurations during testing as well as in speaker adaptive training scheme. In the former case, we employ VTLN as unsupervised for each testing utterance without the knowledge of transcription of adaptation data. Our results on different types of broadcast programs show that the resulting approach for VTLN reduces the Word Error Rate (WER) of our system significantly - by 7 % relatively.