AA: I'm Avi Arditti with Rosanne Skirble, and this week on Wordmaster: an electronic tutoring system that helps non-native speakers of American English learn to pronounce words with a native accent.
RS: The product is called NativeAccent. It's sold by the Carnegie Speech company, with software technology under license from Carnegie Mellon University in Pittsburgh. The system is being used, for example, to train customer service representatives at call centers in India, so they are more understandable to native English speakers in other countries.
AA: We talked to Gary Pelton, director of product development for Carnegie Speech, and Jaime Carbonell, director of the Language Technologies Institute at Carnegie Mellon. They say the aim is accent to reduce an accent, not eliminate it. We asked Jaime Carbonnell which languages are most difficult for NativeAccent to handle.
JAIME CARBONELL: "The product has models for different source language, different original languages of the speakers. It has models for Chinese, Spanish, French, Arabic, German -- twenty-six different languages. Some of the languages are harder because they are further away from English. Chinese is one of the harder ones, for example."
AA: "So the success rate for NativeAccent, for the software, when it's actually correcting the students, might be lower than if it were, say, Spanish or French?"
JAIME CARBONELL: "It's usually measured as time on task. If you're in a language that diverges more from English, you have to use the product longer before you can get to the same level of performance."
RS: "I just want to get a clear understanding of how this works. Either you or Gary can tell us how a student would approach this machine."
JAMIE CARBONELL: Well, Gary is the expert on that so I'll turn it over to him.
GARY PELTON: "Probably the easiest thing to do is for me to play an example. What I have here is I have a Mandarin speaker that is reading a sentence and the sentence is, 'It's just like the picture in my geography book.'"
CHINESE SPEAKER: "It's just like the picture in my geography book."
GARY PELTON: "So what comes up on the screen is you see the actual words, but some of them are marked in red. What we've done is we compared his speech for that little part in the end of the 'geography,' the little y sound, to how a group of native speakers said geography. And, we noticed that he was far away from that group of native English speakers. He also was far away for the 'eo' sound in geography."
AA: "And, how was the correction suggested?"
GARY PELTON: "So below that, what it does is it looks at the part that is farthest away and then gives him some text to tell him how to make the 'e' sound in geography better and also gives him pictures and so forth of what he has to do in his mouth."
AA: "So does the student hear a correct version too, in addition to the text and the pictures of what to do with his mouth, does he actually get to hear it and replay the correct way to say it?"
GARY PELTON: "Yes. So he actually gets a choice of several different model speakers. Since it is a male, he gets male model speakers. We've done studies that show that that makes a big difference if you're trying to mimic somebody in the same vocal range and speed that what you're working with."
RS: "So is what you are doing with all this text and the correction, is that meant to direct the students so that they become aware that this sound exists?"
JAMIE CARBONELL: "Yes, the original inventor of this idea, her name is Maxine Eskenazi -- she is the one who designed the techniques. And one of the techniques is to grossly exaggerate the difference between the sound that the student is using and the sound that they really should be saying, so that they can hear that there is indeed a difference. And then after they perceive that difference, then you minimize the exaggeration until it comes gradually back to normal."
RS: Jaime Carbonell, director of the Language Technologies Institute at Carnegie Mellon University in Pittsburgh. We also spoke with Gary Pelton, director of product development for Carnegie Speech.
AA: The product NativeAccent was displayed this week at Interspeech 2006: the Ninth International Conference on Spoken Language Processing, at Carnegie Mellon. The conference attracted 1,000 scientists from around the world to explore ways in which people and computers use and understand each other.
RS: And, that's Wordmaster for this week. Our e-mail address is [email protected]. And it's it American English you want to hear, check out our Web site at voanews.com/wordmaster. With Avi Arditti, I'm Rosanne Skirble. |