Context Dependency in Neural Network Based Acoustic Models

Our recent experiments with Gaussian mixture (GMM) based acoustic models have shown that employing context dependent acoustic models, namely triphones, can greatly improve recognition accuracy [1] in comparison to systems based on context independent units. Significant portion of our research has been aimed at exploring the possibilities of neural networks as acoustic models for speech recognition. We have observed, that a neural network can lead to similar recognition accuracy as a GMM acoustic model, while having less trainable parameters [2]. Neural network acoustic model have other interesting properties, such as context sensitivity, the ability to use several subsequent speech frames as an input and thus providing some context information [3].