Industry coverage, news, original feature articles, how to guides, videos, podcasts and reviews of BlackBerry service providers, software, accessories, hosted exchange providers and more

Earlier this month, Yahoo! unveiled its new oneSearch, which includes voice-based search. Since this was an exclusive BlackBerry download, I decided to give it a test run. As I reported at the time, the feature worked well even in loud, crowded areas. For the most part, oneSearch was picking up my voice with accuracy, though there are clearly nuances in my speech that tripped it up. But as we were later told by Yahoo! execs, oneSearch will learn my style of speech over time. So I’m here to check back in 28 days later.

Searching for acronyms

As I noted earlier, the biggest issue seemed to be oneSearch’s recognition of individual letters. Most notably, it was confusing “c” and “e,” “m” and “n,” and other like-sounding letters. So that’s the first issue I set out to settle.

The search was for CPM — cost per mile, or cost per mille (in the advertising world). I tried my best to use a consistent inflection on each of my searches, as to not purposely trip up the voice recognition software.

First try: oneSearch picked up “cynthia.” Clearly, we’re not off to a good start. So I went up to the search bar and typed in “CPM” manually. Let’s see how it helped.

Second try: it heard me say “CPN.” Not bad. Just one letter off, which I promptly corrected in the search bar.

Third try: truly a charm. It returned “CPM.”

A further experiment

Now, for a bit of a change. Now that it’s used to hearing me say “CPM,” I wonder what it will do when I say “TPM,” as in Talking Points Memo.

It came up with “PPS,” which is clearly way off. So is oneSearch just learning specific search terms, rather than the general things I’m saying? It appears so. But let’s see if it continues to learn.

Second try: “TTM” Getting closer. Will the third time bring victory this time around?

Third try: “TTM” again. Now I’m starting to think I’m not speaking slowly or clearly enough. I corrected the search, and will give it a fourth try — though I’m not optimistic, given the identical interpretations on the second and third tries.

Fourth try: “TTM” yet again. So it appears that it’s not learning from my corrections.

I wonder how well it remembers my previous queries. And whaddya know? It returned “CPM” perfectly on the first try.

Speaking slowly

My guess now is that oneSearch simply can’t separate letters well enough when speaking at a normal rate. So let’s try the “TPM” experiment again, since I’d like to see oneSearch succeed here.

I spoke the letters TPM into oneSearch slowly and deliberately into my BlackBerry. Unfortunately, it didn’t like that either. The returned search string: “pee pee in my.” No joke. A second identical search yielded “p e t m.” A third search gave me “c t and.” For a fourth search, I sped up a bit, and was back to “TTM.”

I tried it a few more times, trying to vary my inflection to make it very clear that I was searching for the letters t, p, and m in succession. Every freakin’ time I got “TTM.” Speaking slowly and speaking quickly, I was not able to get oneSearch to recognize this particular string.

Strange names

There is a pitcher for the Chicago White Sox named Mark Buehrle. It’s pronounced like “burly.” So I wonder what comes up when I search for his name.

“Marc Berlin,” was what oneSearch returned, who is apparently a film director. Time to put in my manual correction. Only problem: It doesn’t recognize the name Mark Buehrle. At all. So I’ll try for another major league pitcher, Justin Verlander.

Bingo! First try and I’ve got it. It’s got news, images, and Web results for the fireballing righty. But his name is easy. It’s uncommon yet phonetic, so there’s little room for error. How about Adam Dunn? Will it give me “Adam done”?

“I and dunn” is what it returned. This is where I started to really notice the strengths and weaknesses of oneSearch. Just one more, so I can share my revelation. Search term: “Constantinople.” Perfect on the first try.

More syllables means more accurate search results

It seems to be as simple as that. When searching for strings of letters, I’m using monosyllabic phrases. Apparently, this is no good for oneSearch. While it seemed to learn “CPM” relatively quickly, it could never quite differentiate my “P” from my “T” in “TPM,” constantly returning me “TTM.” This is backed up by my search for “API,” which unsurprisingly returned “ATI.” It makes “CPM” an outlier, I guess.

(A search for “PBS” returned accurately, though, suggesting that Ps in the middle of acronyms are the real problem here.)

But look at the searches for “Verlander” and “Constantiople.” They’re polysyllabic words, and they returned perfectly on the first try. Even the term “the first world war” returned perfectly, even though each word in the string is one syllable.

I’m sure these are issues that improve as we move along with oneSearch. If the recognition of single letters is the only issue at hand — and really, that’s the only trouble I encountered — I think that’s something that’s correctable.

Share on Facebook | Add to De.licio.us | Digg This! | Technorati