| Paper: | SLP-L12.2 | ||
| Session: | Discriminative Training | ||
| Time: | Friday, May 19, 14:20 - 14:40 | ||
| Presentation: | Lecture | ||
| Topic: | Speech and Spoken Language Processing: Discriminative Training Methods | ||
| Title: | Recent Improvement on Maximum Relative Margin Estimation of HMMs for Speech Recognition | ||
| Authors: | Chaojun Liu, Panasonic R&D Company of America, United States; Hui Jiang, York University, Canada; Luca Rigazio, Panasonic R&D Company of America, United States | ||
| Abstract: | Our previous study on maximum relative margin estimation (MRME) of HMM demonstrated its advantage over minimum classification error (MCE) training. Here we report our recent improvement on MRME. Two novel approaches are proposed to handle recognition errors in training set for MRME. One is a new training criterion based on a combination of MRME and MCE objective functions. The other one proposes to remove a strong constraint in the original MRME algorithm, so that MRME can be applied to all training data as opposed to only correctly recognized data in the original MRME. Both approaches can take advantage of more training data during the large margin training and can bootstrap directly from MLE model without a separate MCE training. Improvement on recognition accuracy has been achieved on a connected digit string recognition task using TIDIGITS corpus. | ||
©2018 Conference Management Services, Inc. -||- email: webmaster@icassp2006.org -||- Last updated Friday, August 17, 2012