Hill-Climbing Ensemble Feature Selection with a Larger Ensemble
Title | Hill-Climbing Ensemble Feature Selection with a Larger Ensemble |
Publication Type | Technical Report |
Year of Publication | 2009 |
Authors | Gelbart, D. |
Other Numbers | 2706 |
Abstract | This report follows up on the investigation of hill-climbing ensemble feature selection for multi-stream speech recognition in David Gelbart's Ph.D. thesis. We compare using three streams as in the thesis to using five streams, to investigate the hypothesis that hill-climbing would give larger accuracy gains with more streams due to there being more feature selection options available. To move from three streams to five streams we split the MFCC and PLP streams into separate streams of static and dynamic features. We find that hill-climbing with five streams does result in better word recognition accuracy than three streams. This may be because our hypothesis is correct, but it also might be because the initial five stream systems that hill-climbing started from were more accurate than the initial three stream systems. |
URL | http://www.icsi.berkeley.edu/pubs/techreports/TR-09-001.pdf |
Bibliographic Notes | ICSI Technical Report TR-09-001 |
Abbreviated Authors | D. Gelbart |
ICSI Research Group | Speech |
ICSI Publication Type | Technical Report |