Research pinpoints speech recognition errors

A study carried out by scientists at the University of Edinburgh may help improve automated ticket booking lines by analysing regional dialects.

The research focussed on finding the most frequent speech recognition errors currently made by automated phone systems. It is hoped that, by pinpointing these errors, booking line systems could be developed to understand customers voices better.

The study involved feeding conversations between two people into a speech recognition system and recording how well the speech was understood by the software.

Dr Sharon Goldwater, a lecturer in Informatics, hopes the research can help enhance the current software used: "Voices vary from one person to the next and it is challenging to design a computer system that can understand lots of different voices."

One of the discoveries in the study was that amongst the most frequent faults made by speech recognition services is the inability to recognise the first word spoken in a phrase.

This could be due to the customer inhaling before beginning the phrase, which would confuse the system. Another potential explanation is that the recognition system cannot put the first word into context as it has no other information on the phrase.

Other errors include the lack of ability to recognise speech sounds other than words, such as ‘umm’ and ‘err’. Although the system can distinguish these sounds, they cannot identify them for what they are and instead try to translate them as meaningful words.

Furthermore, the study shows that gender affects the accuracy of the results. Due to their tendency to hesitate or mumble more frequently, men’s speech is more commonly misunderstood than women’s.

Variations in tone, speech and enunciation can also influence the outcome of the readings. By closely studying these differences, scientists could work towards developing systems that can understand different accents.