Note: Your custom schedule will not be saved unless you create a new account or login to an existing account.
| Paper: | SP-P14.12 |
| Session: | Acoustic Modeling: Tone, Prosody, and Features |
| Time: | Thursday, May 20, 15:30 - 17:30 |
| Presentation: |
Poster |
| Topic: |
Speech Processing: Acoustic Modeling for Speech Recognition |
| Title: |
MINIMUM CLASSIFICATION ERROR TRAINING OF LANDMARK MODELS FOR REAL-TIME CONTINUOUS SPEECH RECOGNITION |
| Authors: |
Erik McDermott; NTT Corporation | | |
| | Timothy Hazen; Massachusetts Institute of Technology | | |
| Abstract: |
Though many studies have confirmed the effectiveness of the MinimumClassification Error (MCE) framework for discriminative training ofHMMs applied to speech recognition, few if any have reported MCEresults for large (> 100 hours) training sets in the context ofreal-world, continuous speech recognition. Here we report substantial gains in performance for the MIT JUPITER weather information task as aresult of MCE-based optimization of acoustic models. Investigation ofword error rate vs. computation time showed that small MCE modelssignificantly outperform the Maximum Likelihood (ML) baseline at allpoints of equal computation time, resulting in up to 20% word errorrate reduction for in-vocabulary utterances. The overall MCE lossfunction was minimized using Quickprop, a simple but effectivesecond-order optimization method suited to parallelization over largetraining sets. |
| |
| Back | |