Publication Details
All-Pole Modeling for Definition of Speech Features in Aurora3 DSR Task
speech, speech processing, feature extraction ,speech recognittion, PLP, MFCC
All-Pole Modeling for Definition of Speech Features in Aurora3 DSR Task
This contribution investigates various all-pole modeling based feature extraction techniques for robust ASR. Final performances are compared to well-known Mel-frequency cepstral coefficients. At the beginning, frequency independent approach of modeling speech power spectra that increases the performance of ASR system mainly in case of large mismatch between training and testing data is studied. Then, we focus on different types of features that can be extracted from all-pole model to reduce the overall word error rate. Achieved recognition performances show that line spectral frequencies are more suitable parameters for ASR where the input speech is corrupted by different types of real noises than cepstrum based features. In experiments SpeechDat-Car databases used for front-end evaluation of advanced distributed speech recognition (DSR) systems were used.
@INPROCEEDINGS{FITPUB7234, author = "Petr Motl\'{i}\v{c}ek and Jan \v{C}ernock\'{y}", title = "All-Pole Modeling for Definition of Speech Features in Aurora3 DSR Task", pages = "295--300", booktitle = "6th International Conference, TSD 2003 \v{C}esk\'{e} Bud\v{e}jovice, Czech Republic, September 2003 Proceedings", journal = "Lecture Notes in Computer Science", volume = 2003, number = 09, year = 2003, location = "\v{C}esk\'{e} Bud\v{e}jovice, CZ", publisher = "University of West Bohemia in Pilsen", ISBN = "3-540-20024-X", ISSN = "0302-9743", language = "english", url = "https://www.fit.vut.cz/research/publication/7234" }