| ______ |
What makes a voice application simple? Is it fewer commands?
Actually the tiniest recognition grammars can be quite problematic. "Yes/No" grammars carry a heavy burden of required accuracy compounded by the fact that humans often relax their pronunciation in just such situations. Also as the vocabulary size decreases, the cognitive load on the human is often higher because the human must recall very specific utterances and their attendant implications on the dialog state. This kind of speech interface is actually at odds with our innate language skills.
Vocabulary size (number of possible utterances) impacts the number of conversational turns taken for a given task.
"Yes/No" recognition (the simplest binary system) leads to a game of "twenty questions" that requires the human to keep the objective as well as the history of the conversational states in memory. On the other hand, there is a limit to the size of a task that can be done in one utterance (the most direct system we could hope for).
Once this technology and the deployed infrastructure attains a
level that can support conversational behavior, then there will be a major shift
in the user population and the modes of deployment.
All the current "pre-conversational" systems with their unnatural cognitive burdens (vocabulary, state awareness, etc.) and their need for training and practice will be overshadowed by systems that "fit" our inborn language skills. These early systems may not be immediately recognized since they will start simple. But to the untrained ear and brain they will seem easier to grasp. The systems that succeed will co-opt the human rules of conversation as opposed to arbitrary dialog rules that are designed.
Is it even possible to design a natural system? Or do we need to approach the problem from a fresh point of view?
|
