RealTime IT News

Voice Search Faces Hurdles in Mobile

NEW YORK -- Search is a huge and evolving business. Facebook, not a traditional search company, recently said its own search offering is booming. Incumbent search giant Google (NASDAQ: GOOG) recently said that it will move beyond keywords at SES. And at the SpeechTEK conference this week, big search players talked up opportunities in mobile search.

"Companies are excited, eager, and fearful about the revenue opportunity in mobile search," said Microsoft (NASDAQ: MSFT) Senior Researcher, Geoffrey Zweig, who spoke on a voice search panel at SpeechTEK here.

"Nobody wants to be left behind," he added. He noted that companies have a strong business model for traditional Web-based search, but mobile voice search is a different animal.

While many see opportunities, they also see barriers, panelists said. Services are restricted by factors as various as noise conditions and the need to limit the vocabulary size of recognition engines, which are also known as recognizers, said moderator Michael Cohen, manager of Google's speech technology group.

Zweig listed numerous ways noise conditions can interfere with an attempted voice search. Ambient noise from cars, restaurants, parties, or televisions can make processing more difficult, as can the diversity of U.S. accents.

"About 30 percent of stored speech is repeated," he said, which means that systems frequently don't get it right the first time.

The importance of the first time

Getting it right immediately is very important, panelists said. "If an app doesn't work immediately, then on day two, users won't come back," said Johan Schalkwyk, manager of Google's mobile speech team.

Schalkwyk added that when Google launched its mobile search engine, it was optimized for U.S. English even though it was available worldwide. Google learned that users in Australia and the UK, who received lower quality results because the system was not designed for them, were less likely to use the service again.

Microsoft's Zweig had even more specific data. "If users are doing an average of 10 percent click through on their voice search, there's only an 18 percent chance they'll stay. In contrast, if every search is working and they're doing an average of 90 percent to 100 percent click through, then there's over a 70 percent chance that they'll become a long term user.

Another consideration is that users' expectations need to be managed.

"User retention rates correlate with how well people know the system, such as whether they know that shorter queries work better" said Vlingo speech scientist, Han Shu.

"A user who expects everything will be disappointed," he added.