(FROM THE WALL STREET JOURNAL 5/29/15)
Speech Recognition
Gets Conversational
Smartphones are good at understanding what users say to them,
but they can't handle conversations. The shift from one speaking
voice to another confuses virtually all speech-recognition
software.
Researchers at International Business Machines Corp. say they
have succeeded in building an algorithm that doesn't get tripped up
so easily. The new software accurately recognizes conversations
spoken by two alternating voices, said Michael Picheny, the leader
of IBM's speech team.
The researchers tested the software, which like many IBM
artificial-intelligence efforts will carry the Watson brand name,
on a database of telephone conversations between strangers and
found an 8% error rate, 36% better than the best previous results
on the same test, Mr. Picheny said.
That's not as accurate as a person, but it's close.
"Humans on this particular set of data only get 4% of the words
wrong," he said. "A few years ago the number was closer to 20%" for
the best software available.
Even Facebook Inc., one of IBM's rivals in the
artificial-intelligence race, is impressed. The results represent a
"significant advance," said Yann LeCun, Facebook's director of
artificial intelligence research.
The success of the IBM team stems from fusing a variety of
so-called deep-learning computing techniques. It used deep neural
networks to spot speech patterns; convolutional neural networks,
also called ConvNets, to sort out variations in pitch; and
recurrent neural networks to remember past sounds.
Many researchers lately have been combining these techniques. In
the past year, Baidu Inc., Google Inc. and Microsoft Corp. have
published papers on these techniques.
"People at IBM and elsewhere have used ConvNets and neural
language models before, but this is the first system that really
puts it all together and beats a long-held record on a widely
accepted benchmark," said Mr. LeCun, one of the inventors of the
ConvNet technique.
The new software will be added to a programming service, known
as an application programming interface, or API, released in
February, that lets programmers use Watson for speech recognition,
Mr. Picheny said. It could be used, for instance, to help
customer-support staff with callers' whose speech is difficult to
understand.
The artificial-intelligence techniques involved have been around
since the 1980s, but they have become hot areas of applied research
and development only in the past five years.
Google has spent hundreds of millions of dollars acquiring
companies with neural networking expertise. Apple Inc., Baidu,
Facebook, IBM and Microsoft also have been snapping up
deep-learning experts in an effort to build systems that recognize
speech, photographs and videos with near-human accuracy.
-- Robert McMillan
Splunk's Loss Widens
Despite Solid Revenue
Big-data software company Splunk Inc. on Thursday reported
another period of strong revenue growth, though higher expenses led
to a wider loss for its fiscal first quarter.
For the second time in as many quarters, the company increased
its revenue guidance for the year.
Splunk said it had a strong quarter and that its 9,500 customers
now include 80 of the Fortune 100. "We welcomed a record number of
new customers to Splunk Cloud," the company said.
For the year ending in January 2016, the company now expects
revenue of $610 million to $614 million, compared with a previous
view of $600 million.
In the current quarter, it expects $138 million to $140 million.
Analysts polled by Thomson Reuters expected second-quarter revenue
of $136 million.
While Splunk's revenue and customer base have been growing
rapidly, the company has yet to post a profit.
Earlier this month, Splunk hired Snehal Antani as chief
technology officer. He replaced Todd Papaioannou, who left the
company in November.
For the quarter ended April 30, Splunk reported a loss of $71.2
million, or 57 cents a share, compared with a year-earlier loss of
$50.8 million, or 43 cents a share.
Excluding stock-based compensation and other items, the company
reported a loss of one cent a share. Analysts surveyed by Thomson
Reuters expected a loss of three cents a share.
Revenue jumped 46% to $125.7 million. The San Francisco company
had projected revenue of $116 million to $118 million.
-- Josh Beckerman
Apple Buys Metaio,
German Reality Firm
Apple Inc. has acquired German augmented-reality firm
Metaio.
Apple confirmed the acquisition with its standard response when
it buys a company. "Apple buys smaller technology companies from
time to time and we generally do not discuss our purpose or plans,"
said an Apple spokesman.
Metaio has worked on a technology that allows people wearing
Google Glass-like eyewear to make any real-world surface into a
virtual touch screen. It also created software to let developers
create tablet apps that flash repair instructions across the screen
when the tablet's camera is pointed at machines such as car engines
or a printer.
The acquisition comes a few months after Piper Jaffray senior
research analyst Gene Munster said Apple has a small team exploring
augmented reality technology. He said that Apple could help develop
AR products that appeal more to consumers than products like Google
Glass.
Apple news site 9to5mac reported earlier this week that Apple is
working on adding augmented- reality features to its Maps
application. The feature would allow a user to point an iPhone
camera down a street and it would overlay information about
restaurants, cafes and other local businesses in that vicinity.
The acquisition was earlier reported by 9to5mac.
-- Daisuke Wakabayashi
Access Investor Kit for Apple, Inc.
Visit
http://www.companyspotlight.com/partner?cp_code=P479&isin=US0378331005
Access Investor Kit for International Business Machines
Corp.
Visit
http://www.companyspotlight.com/partner?cp_code=P479&isin=US4592001014
Subscribe to WSJ: http://online.wsj.com?mod=djnwires