8th International Conference on Spoken Language Processing

Jeju Island, Korea
October 4-8, 2004

TRAP based Features for LVCSR of Meting Data

Frantisek Grezl, Martin Karafiat, Jan Cernocky

Brno University of Technology, Czech Republic

This paper describes using temporal patterns (TRAPs) feature extraction in large vocabulary continuous speech recognition (LVCSR) of meeting data. Frequency differentiation and local operators are applied to critical-band speech spectrum. Tests are performed with HMM recognizer on ICSI meetings database. We show that TRAP features in with standard ones lead to improvement of word-error rate (WER).

Full Paper

Bibliographic reference.  Grezl, Frantisek / Karafiat, Martin / Cernocky, Jan (2004): "TRAP based features for LVCSR of meting data", In INTERSPEECH-2004, 437-440.