|
Search engines hunt for sound
Friday, May 28 2004
To be Googled, or not to be Googled?
That was the question facing National Public Radio's on-line director
Maria Thomas earlier this year. The answer would seem obvious for
anyone doing business on the Web, where being included in search
results can mean the difference between success and oblivion.
But in Ms. Thomas' case it wasn't that simple: Most "spiders"
that crawl and index the Web are effectively blind to audio and
video content, making NPR's highly regarded radio programming all
but invisible to mainstream search engines.
Indexing files by looking at their audio features is still a work
in progress for big search engines, including Google. So NPR eventually
hit on a plan to instantly turn audio broadcasts into text files
that can be recognized and picked up by search engine spiders.
"Our site is primarily full of rich audio, and we want people
to find it when it's relevant," Ms. Thomas said. "The
big search engines' technologies don't have the ability to get inside
the audio or video. With the little bit of text we have on NPR,
it's not always good enough to find our content, and reference the
page."
Consumers armed with broadband connections at home are driving
new demand for multimedia content and setting off a new wave of
technology development among search engine companies eager to extend
their empires from the static world of text to the dynamic realm
of video and audio.
The stakes are enormous, not just for the search engines, but for
content owners hoping to harness the Internet, stand out in the
on-line information glut and attract new audiences. The winning
search companies could become the gatekeepers in a new era of media
increasingly defined by consumers' ability to seek out programming
on their own terms and consume when and how they want.
NPR is not the first company to bend over backward for recognition
on the search engines, which drive the bulk of traffic to hundreds
of thousands of Web sites. But it may be the first broadcaster to
adopt a guerrilla-like strategy for insinuating its audio in the
search indices. And the company's strategy is working so far: In
recent weeks, NPR audio has begun regularly appearing on the index
pages of Google News and Yahoo News, and clips also crop up when
people search for news-related keywords, such as "Abu Ghraib,"
the name of the notorious prison in Iraq.
Ms. Thomas did not divulge specific traffic figures, but she said
that since stories started showing up in results churned out by
major search engines, the NPR site has seen record spikes in visitors
for high-interest news stories, such as the murder of U.S. hostage
Nick Berg in Iraq.
NPR's move points to the limitations of Google and Yahoo at a time
when broadband Internet connections are becoming more popular among
consumers, fostering new demand for multimedia content. Publishers
are increasingly adding exclusive audio and video content for on-line
access; educators are streaming courses on-line; broadcasters are
bringing vast archives on-line in digital form; and small-time publishers
are finding it cheaper and cheaper to create, produce and host multimedia
on the Web. Yet you wouldn't find any of it on the primary search
engines.
That's created an opportunity for specialty search engines focused
on filling the gap. Already, technology companies including Singingfish,
StreamSage, Hewlett-Packard, Virage, Nexidia and others have emerged
to address some of the challenges. Yahoo and AOL are players, too.
Yahoo owns AltaVista, which has one of the Web's oldest audio and
video search engines, but so far, Yahoo has not sought to feature
the technology. America On-line, another dark horse in the search
race, bought Singingfish earlier this year.
"There's a tremendous upsurge in the amount of streams available
on the Internet," Singingfish General Manager Karen Howe said.
"Because of broadband adoption, which is so strong in the enterprise
and in the home, accessing high-quality content now is not such
a pain for the user." Singingfish fields about six million
searches a day, up from three million in January, and records about
80,000 new streams a day.
Yahoo's and Google's drawbacks could underscore the need for even
more specialty search engines or prompt advancements from the status
quo. Unleashing new features almost daily, the major search engines
are in a race to win the hearts and clicks of Web surfers with their
search tools, largely because it means more advertising revenue
in their pockets. And with Google's upcoming $2.7-billion (U.S.)
IPO, the competition could get even more heated, launching the two
companies into new realms of rivalry.
The technologies that Yahoo and Google rely on today are focused
on mining text on the Internet for content relative to keywords.
Among other techniques, they analyze the interconnectedness between
Web pages and examine the headers and anchor text on a page so that
they can return appropriate Web sites for any keyword, or set of
keywords a surfer does a search on. Even images — an increasingly
popular source for Web searches on Google and Yahoo — are tied
to text that defines the pictures.
Similarly, some search engines now analyze text or keywords that
describe a multimedia file's content, or what's called metadata.
Singingfish, for example, relies on 70 fields of descriptive material
about a file — such as author, bit rate, file size — to
catalogue it, but the company regularly runs into shortcomings in
such information.
Others transcribe portions of the audio or video and then analyze
the language for meaning, topics of conversation, and relevance
to a search term.
Most ambitiously of all, a few are bent on searching inside the
files to extract meaning and relevance by examining audio and video
features directly.
StreamSage is starting to make waves with its audio and video search
technology, introduced late last year. The company, based in Washington,
DC, developed software after roughly three years of research that
uses speech recognition technology to transcribe audio and video.
It then uses contextual analysis to understand the language and
parse the themes of the content. As a result, it can generate a
kind of table of contents for the topics discussed in the files.
The downfalls of this method are that it can be extremely difficult
to be 100 per cent accurate. In fact, experts say the language-detection
technology is typically only 80 per cent accurate. Language hurdles
such as accents, jargon and dialect can trip up the technology,
for example.
StreamSage introduced a Web site called CampaignSearch.com last
week to showcase its technology. The site lets people search audio
and video files on the Web for clips from the presidential candidates,
including files on sites such as Whitehouse.org, CSPAN Voice of
America and others.
"It's a timely demonstration of StreamSage's technology,"
said Seth Murray, the company's president.
For example, the technology can dissect an hour-long speech from
Democratic presidential nominee John Kerry to find a segment in
which he talks about health care, and earmark that 4 minutes of
the broadcast for access.
StreamSage has flown under the radar during its four years of operation
while it has invested heavily in research and development. Its chief
scientist, Tim Sibley, is known for his work in computational linguistics.
StreamSage has received funding from research grants, including
the Massachusetts Institute of Technology's Advanced Technology
Program. Harvard University uses StreamSage's technology to allow
medical school students to search past lectures on related subjects.
AOL is using the technology to provide closed captions for streaming
video and audio on AOL Broadband.
NPR is using StreamSage to transcribe its audio programs as they're
broadcast, thereby helping them to get listed faster. NPR does commission
transcripts for many of its programs, but the traditionally manual
process of transcription would be too slow for a search related
to timely news. Using speech recognition technology, StreamSage
can create text from audio much more quickly, and then feed those
transcripts to Google and Yahoo.
NPR's Ms. Thomas said her outfit eventually replaces the transcripts
from StreamSage with those from humans because the human-rendered
records much more accurately reflect the audio and video content.
StreamSage's results can be garbled.
NPR also licenses technology from Singingfish to meticulously label
its audio files with relevant information, or metadata.
In its own first step toward offering multimedia search, Google,
based in Mountain View, Calif., registers NPR audio files on Google
News, its specialty news aggregation service. A search for a headline
topic that is discussed on audio-only NPR programs, such as Talk
of the Nation, will uncover a link to the audio program and the
specific segment covered, for example.
A Google representative confirmed a relationship with NPR but declined
to comment further on the technology. Up until now, Google has not
listed multimedia files because the company has sought to avoid
the legal uncertainties of indexing and linking to copyrighted works
that owners may want protected, company executives have said in
the past. Beyond those reasons, audio and video file searching can
be a much more difficult technical task to solve than cataloging
the Web.
StreamSage's Murray said he's not worried about potential copyright
issues because his company is not housing the information. Rather
Streamsage points people to audio and video around the Web, just
like Google or Yahoo does.
Exactly how far search engines can go in linking to multimedia
files has yet to be worked out definitively in the courts. The recording
industry last year quietly settled a long running dispute with MP3Board.com
over alleged illegal links to music files without any money changing
hands, said MP3Board's attorney, Ira Rothken.
Yahoo also announced a relationship with NPR, in February, when
it outlined its "content acquisition" program, a systematic
effort to include more hard-to-get information in its searchable
database.
An NPR affiliate in Boston, WBUR.org, is using similar technology
from Hewlett-Packard, called Speechbot. Robin Lubbock, director
of new media for WBUR, said the broadcaster is using Speechbot to
translate audio into text so employees and visitors can search for
content on its own Web site.
Virage, which is now owned by Autonomy, has technology that analyzes
in-stream audio and video and lets people zip to the part of the
stream they want. Yet it can be an expensive enterprise solution.
Jay Webster, chief technology officer of interactive agency Fathom
On-line, said that for most audio or video broadcasters to get ranked
in search engine results, they would have to employ some manual
indexing of their own first.
"Where it gets cool is if you could search on any keyword
and find it within audio and that audio would come up in search
results," Mr. Webster said. "But I don't think we're there
yet."
|
