How Do You Measure Accuracy in Speech Recognition?

Automatic TRANSCRIPT

Do you mean when you say accuracy when it comes to speech recognition. So that's a really really good question. How do we measure that when we talk about accuracy we're talking about. How many words correct did the system get so. We call this word race so often. It's you know the smaller. The number the more accurate system or we can talk in terms of accuracy. The higher the number the more accurate unperformed system is on what happens is as children get younger. The systems that have been on adult data. They have been modeling language adult behaviors. They work reasonably well for amateur ten-year-old nine ten ten twelve the beginning to sound like little adults now we all know that not all kids are like some kids take a little more development maturing but by the whole when they start getting a little bit older and they start looking adults. The system start form. Well what happens when children younger. What you see is the performance. The accuracy error rates of how many mistakes the system gets starts to die right. There's two different types of areas we talk and when you look at it is simply that the system isn't hearing the difference between a child and adult speech or is it in the language interpreting so sometimes when you talk to a voice assistant how you may be asking for something way out there that the system is no idea. So is able to transcribe what you said but it just can't match what you said to any reasonable response you get account response gone. I can't help you right now. I don't know that answer. Canned responses are quite a thing in speech recognition league. We're all aware of but there's another type of problem that exists is that when people with accents or dialects that are not served by these speech recognition systems. That you're starting to see some acoustically. It actually can't even transcribe that. I'm really see that problem when it comes to kids too. We like to think much more as if you're trying to deal with a system to understand a fifty year old male at a four year old girl you can start to see how different the child's speeches their language behaviors whether you're also trying to do for the pheromone your effectively boiling the ocean. You know it's going to work for some courts better than others and that can be age profiles. It can be in language dialects. And that's where we start to see bias so the system. Not performing for certain demographics as well as it does for others and dash can have very profound effects

Coming up next