Search and Co-Opt

There are two ways to confront the pirating of copyrighted material on the Web. One, pioneered by the music industry and embraced now by Hollywood, throws lawyers in the path of digital progress. Every month, NBC Universal demands that YouTube remove snippets from some of the network's most popular shows. Likewise, Viacom, whose cable properties include MTV and Comedy Central, recently filed a $1 billion lawsuit against YouTube for "massive copyright infringement." That may keep the lawyers fat and happy, but it doesn't accomplish much else.

The other approach, conceived by PodZinger, a video-search startup in Cambridge, Massachusetts, is this: Co-opt the pirates. Unleash them to spread your media virally, and let PodZinger track viewership--and kick back ad revenue. That makes "piracy" profitable to the copyright holder.

At the core of PodZinger's proposed solution is video search, a problem it has largely cracked. (Other upstarts such as Blinkx have too, though they tackle it from different angles.) Harnessing 30 years' worth of government-funded R&D in speech recognition--the company is a spin-off of BBN Technologies, a high-tech military contractor that helped create ARPANET, forerunner of the Internet--PodZinger spiders the Web looking for videos and dissects RSS feeds for updates. When it finds a match, it uses voice recognition (juiced by algorithms known as "hidden Markov models" that bet on the probability of a word given its pronunciation and grammatical context) to create a rough transcript of the audio, then classifies the content by topic.

That's vastly different from Google and Yahoo's approach. They simply scan a video's metatags, the words that describe a video file. Although PodZinger's transcripts are currently only 70% accurate, its approach has the potential to transform the search business. For users, PodZinger's Web site offers the ability--finally--to plug in a search term, then skim the results as they would text; click on a word, and they're taken to that exact place in the video. Of course, with billions of Web pages, PodZinger hasn't come close to ferreting out everywhere videos lurk, but its reach is growing: On YouTube alone, PodZinger transcribes some 20,000 new posts each day. For advertisers (there are only a few at this early stage), the company has copied a page out of Google's playbook, offering the video equivalent of keyword ads based on what users search for.

But the real revolution might be for the copyright holder: PodZinger's spiders will in time be able to track down specific video content on command--a clip from last night's Daily Show, for example, or everything that belongs to Comedy Central--and insert an ad into each segment, no matter where it is playing. In other words, PodZinger could force each and every YouTuber to watch a short commercial if they want to see the clip they asked for, then tally the number of times it's played so the advertiser could pay the copyright holder directly. And what if the person posting the material doesn't want the ad? Tough luck; it's not his video.

In essence, PodZinger wants to make allies of what are now two opposing parties. "Bootlegging is going to happen anyway," says Alex Laats, PodZinger's CEO. "Why not make money in a reasonable way? If people can get paid for their content, and you can track when it is viewed, and advertisements can deliver their brand message, then who cares?" As Laats sees it, this way, everyone stands to benefit from the video boom: The copyright holders, the pirates, the fans, and PodZinger, which would skim a few cents off the top.

The man behind PodZinger's speech recognition is BBN chief scientist John Makhoul, who is originally from Lebanon and received a PhD in electrical engineering from MIT. Back in the 1970s, Makhoul and his team started with 50 words, mostly numbers. It took a decade before a computer could string these 50 words together, deciding that a word was, given its context and phonetic pronunciation, the mot juste. Now Makhoul has developed a tool that helps intelligence analysts scour foreign television broadcasts in Chinese, Arabic, and Spanish and translate them into English. The software can even identify a speaker's unique speech characteristics so that, for example, Osama bin Laden tapes aired on Al Jazeera can be instantly tagged. Of course, it's not perfect. Depending on the speaker's accent, "Iraq" can end up "a rock," "in person" can be rendered "in prison," and "how to light for portraiture" can become "how to light for torture."