BROWSE
Detail
Conference Paper
Using Confidence Vector in Multi-Stage Speech Recognition
Cited - time in 
Share 
Share
- Authors
- Hyung-Bae Jeon, Kyu Woong Hwang, Hoon Chung, Seung Hi Kim, Jun Park, Yun Keun Lee
- Issue Date
- 2008-01
- Citation
- International Joint Conference on Natural Language Processing (IJCNLP) 2008, pp.1-5
- Language
- English
- Type
- Conference Paper
- Abstract
- This paper presents a new method of using confidence vector as an intermediate input feature for the multi-stage based speech recognition. The multi-stage based speech recognition is a method to reduce the computational complexity of the decoding procedure and thus accomplish faster speech recognition. In the multi-stage speech recognition, however, the error of previous stage is transferred to the next stage. This tends to cause the deterioration of the overall performance. We focus on improving the accuracy by introducing confidence vector instead of phoneme which typically used as an intermediate feature between the acoustic decoder and the lexical decoder, the first two stages in the multi-stage speech recognition. The experimental results show up to 16.4% error reduction rate(ERR) of word accuracy for 220k Korean Point-of-Interest (POI) domain and 29.6% ERR of word accuracy for hotel reservation dialog domain.
- KSP Keywords
- Computational complexity, Error reduction, Intermediate input, Multi-stage, Overall performance, Point of interest, Reduction rate, Word accuracy, new method, speech recognition