When Not Understood

We have built in mechanisms for coping with people not understanding us. These include actions like speaking louder, more slowly, waving our arms to try to draw a picture of what we’re describing, and repeating our words.

When we’re developing voice interactive devices, there are a few waves we can assess that we’re not understanding what the user is telling us:

• STT only returns partial results or has a low confidence score
• The user asks for the same request two times in a row with a show period of time
• The device is triggered two times in a row with one of those times being a null response

More advanced ways of detecting that we’re not understanding a user include:

• An increase in speech volume
• A change in pitch
• Adding an question inflection at the end of the speech
• Watching for a particular phrase that connotes misunderstanding (such as speaking the trigger word in the command)

When we know we’re not understanding a command, there are three ways we can address this:

1. Request that the user restate the command (easy)
2. Ignore the command altogether and wait for the user to restate the request (easy but can be annoying)
3. Try to ask for clarification around the domain the user is asking (e.g. “At what time would you like to be woken up?” — more difficult to pull off)