A pair of South Korean electrical engineers have worked out a new porn filter that analyzes audio for tell-tale signs of things you really shouldn’t be watching at work. By using sound, they avoid the problem of visual porn-identifiers (pornifiers) that can get tricked by any expanse of skin, as in closeups of the face or other not-inappropriate body parts.
The engineers started by making spectrograms, basically a visual representation of a sound clip, of lots of different kinds of audio, including music, non-porn video, and porn. By analyzing these spectrograms, they figured out that pornographic audio has a few unique qualities that make it fairly easy to recognize: Regular speech is low-pitched, music has lots of different pitches, and both tend to be fairly constant. But the porn spectrograms showed a high pitch that changes often and also repeats itself.
Using this data, the engineers developed software that could identify porn correctly about 93% of the time. There were some hiccups; apparently laugh-tracked sitcoms sound like porn (weird fact!) and porn can sometimes sneak by the censors by using background music. Of course, image-recognition porn identifiers are just about as accurate and require much less time to analyze (a single frame, versus the audio identifier’s need for a longer clip). But it doesn’t necessarily need to be one or the other: Some see a potential to combine the audio and visual identifiers into one super-detector, a gauntlet through which no porn can pass.
