JSALT 2015 Research
Research Group: Structured Computational Network Architectures for Robust ASR
Despite the recent progresses, ASR systems still perform poorly under the far-field low-SNR, reverberation, and multi-speaker conditions. In this project, we attempt to attack these problems by designing novel structured computational network architectures that mimic human being’s behavior of prediction, adaptation, classification, and generation in recognizing speech. We will develop novel factorized adaptation algorithms that consider multiple sources of auxiliary information. The project will be conducted using Kaldi, the leading ASR toolkit, and the computational network toolkit (CNTK), which is designed to train and decode arbitrary computational networks.
- Dong Yu (Microsoft Research)
- Khe Chai Sim (National University of Singapore)
- Liang Lu (University of Edinburgh)