Arnab Ghoshal research description

I am broadly interested in statistical modeling techniques for automatic speech recognition. Over the past few years I have worked on acoustic models that factorize different sources of variation. Such models may be easily adaptable across diverse domains by factorizing out confounding covariates. Specifically, I have worked on developing the algorithms and tools for speech recognition with the subspace Gaussian mixture models and studied its application to multilingual and cross-lingual speech recognition. More recently, I have started working on deep neural networks for speech recognition. We have shown that the hidden layers of a deep network, obtained from either unsupervised pretraining or supervised finetuning, are transferable between languages making such models suitable for application across diverse domains.