New MS Office nears the hands-free sweet spot

By MICHAEL FOREMAN

Speech recognition technology has been bubbling just under the surface for many years, but its inclusion as a standard feature of Microsoft's Office arguably puts it into the mainstream.

In the latest XP version of Office, not only can you dictate text but you can use voice commands to navigate applications like Word or Excel.

But the question is, does it work? After a couple of days' testing, we can report that it does - well, sort of.

Like most speech recognition programs, Office needs to be taught the characteristics of your voice before it will work. This is achieved with a training wizard that displays snippets of text, which you read into a microphone.

Microsoft recommends that you find a quiet place to do this to avoid background noise, but it might also spare you some embarrassment - it's easy to sound like a robot when you're trying to speak clearly without placing any special emphasis on words.

As each word is recognised, it is highlighted, which reassures you that you are being understood. Surprisingly, the training wizard seemed to perform best when you read at a fairly fast speed without pauses between words.

After one training lesson of about 15 minutes, Word was ready to take its first dictation. "The quick brown fox jumped over the lazy stile," I announced self-consciously into the microphone.

A couple of seconds later, I was amazed to see the first five words of this sentence appear on the screen, reproduced faithfully word for word.

But reality soon intervened when the final part of the sentence was rendered a moment later as "under the ladies the stellar."

After another 15 minutes of training, the program recognised the quick brown fox sentence correctly about three out of four times, except for spelling stile "style."

Microsoft claims that users can expect 85 per cent accuracy when they start using the program, and 95 per cent after sufficient training.

At this level of accuracy, speech recognition would indeed approach the sweet spot that has been reached by optical character recognition (OCR), but unfortunately I achieved nowhere near this.

After about an hour of training, I found the program could recognise words correctly only about 60 per cent of the time, with long words being understood more successfully than short ones. But, considering that the speech recognition facilities are designed for North American users, and I have a British accent with kiwi overtones and was suffering from a severe cold to boot, I was mildly impressed.

The frequent mistakes were often amusing. For example, "Microsoft Office XP Standard" was translated as "Microsoft office decks please stand up," and "the New Zealand Herald" came out as "The museum terawatt." I'm still not sure whether this is a reflection of the program or my cold.

At one point I left the program running in dictation mode while a TV was switched on. When I returned a couple of hours later I found a huge Word file had been compiled in my absence, containing pages and pages of text that read like the demented ravings of a lunatic.

The program had managed to pick up the odd word or phrase correctly from a news bulletin, but the preponderance of "oohs" and "aahs" towards the end of the document made me wonder exactly what sort of programme I had missed.