CYBERSPACE—According to doctoral students MyungJong Kim and Hoirin Kim, porn filter developers have been looking in all the wrong places, literally. Like Jodie Foster in Contact, they should have been listening instead, only much closer to home.

The two electrical engineers at the Korea Advanced Institute of Science and Technology in Daejeon, South Korea, are so convinced that using sound to distinguish porn videos from other fare could be more accurate than using the common image-based methods that they have developed a "signal-processing technique called the Radon transform to create spectrograms of a variety of audio clips, each just half a second long."

According to the New Scientist, "The researchers used a statistical model to classify sounds as pornographic or non-pornographic according to their spectral characteristics, and tested it on audio taken from online videos. The non-sexual audio clips included music, movies, news and sport."

Using test clips, the model reportedly performed with a 93 percent success rate, but the false positives noted in the article—background music the system thought was lewd action or laughter in comedy shows mistaken as "loud audience cheers and cries [that] share similar spectral characteristics to sexual sounds—might be an insurmountable weakness in the model as a singular solution. The fact that audio analysis requires longer clips than visual, which can do the job with a single frame, may be another.

New Scientist spoke with Richard Harvey, a computer scientist at the University of East Anglia in Norwich, U.K., who has worked on image-based porn detection. He thought the Korean's idea was "ingenious," but suggested that because image-based methods achieve similar success rates it might make sense to use both visual and aural detection methods for even better results.

The two Kims present their research at an International Workshop on Content-Based Multimedia Indexing in Madrid next month.