Context-Aware Speech Interface for Human-Robot Interaction

Kurzfassung

The focus of this thesis is the emerging field of human-robot interaction (HRI). Speech is the most intuitive way of communication between humans. This thesis aims at introducing speech recognition capabilities to RAZER, an intuitive graphical human-robot interface with an integrated framework. Additionally, the architecture of the framework is extended generically in order to support multiple interaction methods. While interaction via a graphical user interface (GUI) is effective, it may not be suitable for the cases where the user needs their hands to be free. Thus, further interaction methods are needed as an alternative or as complementary to the GUI. In the scope of this thesis, different speech recognition tools and speech recognition limitations are analyzed. A custom language model is created for speech decoding. An I/O service is implemented in order to support different interaction methods. A speech interaction concept is implemented via CMU Pocketsphinx and the custom language model created. For the evaluation of the system, a user study is conducted. The system is evaluated in terms of usefulness and efficiency. Furthermore, Google Speech API and CMU Pocketsphinx tools are compared using both generic and customized language models for Pocketsphinx by means of accuracy. The results suggest that speech interface is a useful complement to the GUI. According to accuracy tests, CMU Pocketsphinx with the customized language model is the most accurate amongst three tools. Google Speech API is the second best tool and CMU Pocketsphinx with the generic language model delivered the worst results.