Guest Lecture: Katrin KIRCHHOFF

Prof. Katrin KIRCHHOFF from University of Washington, Seattle, USA, will present her work 
 
Graph-Based Semi-Supervised Acoustic Modeling in DNN-Based Speech Recognition
on Tuesday, 23.09.2014, 14:00, in our seminar room IDEG134, Inffeldgasse 16c/EG.
 
Abstract:
This talk focuses on the combination of two recent machine learning techniques for acoustic modeling in speech recognition: deep neural networks (DNNs) and graph-based semisupervised learning (SSL). While DNNs are powerful supervised classifiers, graph-based SSL can exploit valuable complementary information derived from the structure inherent in the unlabeled test data. Previous work on graph-based SSL in acoustic modeling has been limited to frame-level classification tasks and has not been compared to, or integrated with, state-ofthe-art DNN/HMM speech recognition systems. In this task I will describe the integration of graph-based SSL into DNN-based speech recognizers and its evaluation on two small-tomedium vocabulary speech recognition tasks. Results demonstrate that the SSL-enriched systems achieve significant improvements in HMM state classification accuracy as well as consistent reductions in word error rate over state-of-the-art DNN/HMM baseline systems. Moreover, graph-based SSL achieves similar word error rate reductions as self-training (another approach that utilizes unlabeled test data) while being computationally more efficient.

 



Date with Time
23. September 2014 - 14:00
Contact
Location
Seminar Room IDEG134, Inffeldgasse 16c/EG