Expectation-Based Command Recognition Off the Shelf: Publicly Reproducible Experiments with Speech Input

When striving for a cheap implementation of command recognition for speech input today, you may resort to off-the-shelf tools. In contrast to specific research approaches, such tools by themselves do not take expectations for certain commands in a given situation into account. Such expectations will usually be available both in "intelligent" and more conventional programs using this speech interface, and they should be used to improve command recognition. We propose to make use of a set of expected commands at a given state of a dialogue and a list of ranked command hypotheses from basic speech recognition. We devised and implemented this with two approaches for speech input. One specializes a given grammar according to expected commands at each dialogue state, the other accepts the highest-ranked hypothesis for a command that fits the expected ones at a given state. The latter approach achieved a statistically significant improvement of the command success rate in an experiment, as compared to ignoring the expectations. Since everything is freely available, we made these experiments publicly reproducible.