Deep Neural Network based Feature Extraction Using Convex-nonnegative Matrix Factorization for Low-resource Speech Recognition

Bottleneck feature (BNF), together with Gaussian mixture models, has achieved great success compared with acoustic features in low-resource speech recognition. However, the existing of BN layer decreases classification accuracy of deep neural networks (DNN). In this paper, we investigate a better way of extracting DNN based low-dimensional features using convex-nonnegative matrix factorization (CNMF). Firstly a DNN is trained without setting the BN layer. Secondly CNMF is applied on the weights matrix of a hidden layer to form a low-dimensional feature extraction layer. Finally a new type of high-level feature is extracted by forward passing input acoustic feature. Experiments show that the new feature produces 1.6-4.6% gain over BNF baseline system in English and Czech low-resource tasks. When dropout and maxout are introduced, 3.1-5.6% additional gain over BNF baseline system is observed while the training time reduces.