MOTIVATION

Continuous, natural speech input to voice user interfaces enables a conversational interaction, but requires sophisticated speech recognition. On the other hand, recognizing isolated keywords requires significantly less training data and computational resources. This could prove really useful in developing countries, where there are a lot of mobile users who obtain critical information over the voice platform.

This paper assesses the usability impact of restricting input to isolated words in a voice-only search interface. To do so, we designed, implemented, and evaluated continuous speech and isolated word versions of Farmers’ Exchange, a voice interface for searching for answers to agricultural questions.

RESEARCH RESULTS

Task completion rates were comparable between the two conditions, but continuous speech was significantly faster. The difference was correlated to longer, more specific initial queries given more often with continuous speech input. However, more input wasn’t always better; there was an “optimal” query length for Farmers’ Exchange. Though a slower interface overall, isolated word was significantly faster for participants with extensive prior experience with search interfaces, suggesting that a longitudinal comparison of the two modes would be a worthwhile future direction.

PROJECT VIDEO

In collaboration with Neil Patel, Steve Marmon, Dr. Scott Klemmer & Dr. Tapan Parikh. As part of HCI Research at Stanford.