http://www.vuidesign.net/four-vui-lessons-from-social-web-design-pitfalls.htm Interface Design Lessons From The World Around Us Fri, 25 Jul 2008 02:33:12 +0000 http://wordpress.org/?v=2.1 http://www.vuidesign.net/four-vui-lessons-from-social-web-design-pitfalls.htm#comment-352 IanRae Fri, 15 Jun 2007 16:16:38 +0000 http://www.vuidesign.net/four-vui-lessons-from-social-web-design-pitfalls.htm#comment-352 #3 is a big issue with speech apps. It's not just that post-deployment tuning is essential. It's the way apps can be failing and no one knows about it. Unless someone is monitoring completion rates, the system can be failing silently. I blogged about this at (http://silentsoftware.blogspot.com/2006/08/you-cant-tune-silence.html), and suggested a temporary tuning mode that adjusts the recognition thresholds to force more confirmations, allowing the app to log the rejected values. #3 is a big issue with speech apps. It’s not just that post-deployment tuning is essential. It’s the way apps can be failing and no one knows about it. Unless someone is monitoring completion rates, the system can be failing silently. I blogged about this at (http://silentsoftware.blogspot.com/2006/08/you-cant-tune-silence.html), and suggested a temporary tuning mode that adjusts the recognition thresholds to force more confirmations, allowing the app to log the rejected values.

]]> http://www.vuidesign.net/four-vui-lessons-from-social-web-design-pitfalls.htm#comment-353 eolvera Fri, 15 Jun 2007 17:45:55 +0000 http://www.vuidesign.net/four-vui-lessons-from-social-web-design-pitfalls.htm#comment-353 Absolutely. But like you said, the current tuning methods aren't very efficient (compared to the almost immediate feedback you get from most website tuning tools) and require a significant amount of human intervention. But one of the things I suggest is maybe looking at other parameters that can be obtained automatically from the system (e.g. In-grammar rates, Accuracy rates, Confidence scores) and creating custom tracking mechanisms to measure whether a certain caller was successful or not at completing a certain tasks, so that the call flows themselves can dynamically adapt based on these results (results which can then be reviewed and analyzed during a full tuning) Absolutely. But like you said, the current tuning methods aren’t very efficient (compared to the almost immediate feedback you get from most website tuning tools) and require a significant amount of human intervention. But one of the things I suggest is maybe looking at other parameters that can be obtained automatically from the system (e.g. In-grammar rates, Accuracy rates, Confidence scores) and creating custom tracking mechanisms to measure whether a certain caller was successful or not at completing a certain tasks, so that the call flows themselves can dynamically adapt based on these results (results which can then be reviewed and analyzed during a full tuning)

]]>