Speech Recognition - Alternative Tokens


Voice recognition -Writing speech recognition engines is a very complex and difficult task. Writing software that makes use of voice recognition engines and libraries is comparatively easier.

One of the issues that you face is that sometimes a spoken word can be matched to several different written words. Let's take an example; the way most people prounounce it is very difficult for a recogniser to tell the difference between 'for', 'fore' and 'four'. That's where alternative tokens come into the picture.

When the engine cannot make an exact match it will give you a set of alternative tokens. In the Java Speech API. You can specify the number of alternatives that you want to see. What the API guarantees is that you will get no more than that number of alternatives. There will be many cases where the engine can make an exact match and alternatives are not needed at all.

Last week I have been playing around with CloudGarden and unfortunately it seems that their is a bug in it. The number of alternatives are not honoured. I still get two or three alternatives even when this setting has a value of one.

Looks like I will have take another look at the IBM implementation. Wish IBM hired a proper web master so that finding information on their site doesn't feel like the quest for karla

comments powered by Disqus