A testbed for voice-based robot control

A hardware and software configuration is being developed to support research and development in speech recognition (SR) as applied to the interactive control of devices. A research goal is to investigate new signal processing techniques and parametric representations such as those used in speech coding to improve communication effectiveness for person-to-person communications, and to apply similar types of processing to improve the performance of machine recognition. SR systems are often developed and trained using prerecorded databases, which often do not capture the range of variation that can be expected in reality. For example, words spoken while reading or in casual conversation sound substantially different in live environments including noise, vibration, distractions, or emotional stress. A live system is desired to set up games or other situations which introduce excitement and competition, with the resulting changes in vocal characteristics. Therefore, the testbed includes high quality recognition software and allows a high degree of interactivity. It is intended to support testing and improvement of speech recognition algorithms and software in a realistic environment. It should be suitable for research (improved audio feature extraction algorithms, model adaptation, etc), engineering (reliable data acquisition front ends), and human factors evaluation. The paper discusses the motivation, chosen software and hardware approach, and some results. I plan to demonstrate the robot responding to various navigation commands using either a conventional microphone or a modified throat microphone.