In this paper recognition results from different speech decoders are presented for Slovenian large vocabulary speech recognition task. For speech recognition two different types of lexica and language models were used. Word based models were used for baseline system and sub-word (stems and endings) based models for comparison. For all decoders a two-pass decoding strategy was used. With all three decoders better recognition results were achieved using stemending models (3% absolute on average). Experiments also showed slightly better recognition results with Julius decoder, as opposed to two other decoders, and improvement of real-time factor for 65%. Introduce Segmeantal Inner Timewarping into
@inproceedings{rotovnik02_icslp, title = {A comparison of HTK, ISIP and julius in slovenian large vocabulary continuous speech recognition}, author = {Tomaz Rotovnik and Mirjam Sepesy Maucec and Bogomir Horvat and Zdravko Kacic}, year = {2002}, booktitle = {7th International Conference on Spoken Language Processing (ICSLP 2002)}, pages = {681--684}, doi = {10.21437/ICSLP.2002-227}, issn = {2958-1796},}Cite as:Rotovnik, T., Maucec, M.S., Horvat, B., Kacic, Z. (2002) A comparison of HTK, ISIP and julius in slovenian large vocabulary continuous speech recognition. Proc. 7th International Conference on Spoken Language Processing (ICSLP 2002), 681-684, doi: 10.21437/ICSLP.2002-227