Effectiveness of phase-corrected rasta for continuous speech recognition

Abstract

Phase-corrected RASTA is a new technique for channel nor m alization that consists o f classical RASTA filtering followed by a phase correction operation. In this manner, the channel bias is as effectively rem oved as w ith classical RASTA, w ith out introducing a left context dependency. The performance o f the phase-corrected RASTA channel norm alization technique w as evaluated for a continuous speech recognition task. U s ing context-independent hidden M arkov models we found that phase-corrected RASTA reduces the best-sentence w ord error rate (W ER) by 23% compared to classical RASTA. For contextdependent models phase-corrected RASTA reduces W ER by 15% compared to classical RASTA.