Researchers engage with phone misery

Published on Thursday 4 March 2010 09:05

IT is a constant source of frustration for anyone trying to do a spot of telephone banking or book a cinema ticket.

But the days of doing battle with a voice-recognition computer may soon be at an end, thanks to Edinburgh University researchers.

They have pinpointed why the high-tech systems are unable to understand certain words.

The study found that computers are regularly flummoxed by human speech when it is peppered with "umms" and "errs", while men are more likely to be the victim of an unsuccessful conversation with technology.

The first word spoken in a phrase is also hard to understand for computers, according to the university, because the machine is unable to put it into the right context. The fact that most people inhale before speaking can also confuse computers.

Dr Sharon Goldwater, of the university's school of informatics, led the research. She said the wide range of accents and pitches made it difficult to create a machine to perfectly understand all voices.

"Voices vary from one person to the next and it is challenging to design a computer system that can understand lots of different voices," she said.

"We hope that by closely studying how people speak and how machines process this, we can help create better systems that will be simple and efficient to use.

"It is a frustrating thing for some people to use, but often if their words aren't recognised the first time they tend to change the way they say things, almost over-pronounce words, or indeed, speak loudly and shout, as if they are talking to a hard-of-hearing person or a foreigner. None of this helps the computer. It is a technology which has improved. Of course it is difficult to improve this kind of thing for a mass audience before it has been exposed to one."

The study, in which Edinburgh University researchers worked closely with colleagues from Stanford University in the United States, was published in the Speech Communication journal, and boosted by support from the US Office of Naval Research.

Words that are not enunciated properly, variations in pitch and a change of speed can also throw computers.

Researchers carried out the study by recording phone discussions between pairs of people and feeding their conversation into a speech recognition system and monitoring how much it could understand.

Ms Goldwater added: "The ultimate objective of this was to help improve the technology."

LISTEN UP

Five tips to succeed with automated speech recognition:

1. Speak slowly, like a radio announcer.

2. Talk without emotion in your voice.

3. Don't over-pronounce words.

4. Don't speak loudly if the first attempt fails.

5. Think before speaking, and leave out "umms" and "errs".

&#149 www.ed.ac.uk