Abstract:

Speech can be an excellent modality for simple information retrieval tasks, such as finding points of interest in an automotive navigation system or selecting music from a personal player. However, to achieve natural yet robust retrieval from spoken inputs in noisy, real-world environments, new techniques are needed. We will demonstrate two such techniques, one software-based and one hardware-based. In the former we use the whole search space of the speech recognition system to perform retrieval; in the latter, we use a new device, an ultrasonic Doppler microphone, to detect the start and end of utterances, thus enabling hands free operation. Together these two techniques can significantly improve the usability of speech interaction for our target applications.