Conclusions and Future Directions
In this last chapter, we draw our conclusions and discuss future directions toward developing a spoken language dialog system.
This book over a solution to enhance the robustness of a statistical automatic speech recognition system by incorporating various additional knowledge sources while keeping the training and recognition effort feasible. A new unied framework has been proposed and the effenciency of its usage has also been analyzed experimentally. A review of our work covering theoretical, application, and experimental issues is given in the following sections.
Despite the rapid progress made in statistical speech recognition, many challenges still need to be overcome before ASR systems can reach their full potential in widespread everyday use.
We learned that only a limited level of success could be achieved by relying solely on statistical models while ignoring additional knowledge sources that may be available. However, there have often been cases where developing complex models and achieving optimal performance have not been simultaneously feasible. This especially applies when there are insucient resources, i.e., the amount of training data and memory space available, for proper training of model parameters. As a result, input-space resolution may be lost due to nonrobust estimates and the increasing number of unseen patterns. Moreover, decoding with large models may also become cumbersome and sometimes even impossible.
KeywordsKnowledge Source Speech Recognition System Noisy Speech Automatic Speech Recognition System Junction Tree
Unable to display preview. Download preview PDF.