DID holiday shopping 鈥� or post-holiday returns 鈥� find you, like me, cursing at a computer over the phone? If so, you were not alone. The of customer-service lines with computerised speech systems spread little joy.
Normally I try to avoid calling, but could find no alternative after a website botched a software order. The voice that greeted me identified itself as a computer, but sounded almost human, and I could hear a keyboard as I tried to explain the problem. For a moment, I was almost fooled into thinking a person was on the line, pretending to be a robot, but realised the simulated typing was too even to be real.
Advertisement
The computer also proved to be dumber than it sounded. It claimed to be looking something up, but eventually admitted it didn鈥檛 understand what I had said. I tried again with no luck, and after another few goes, it became clear that the computer had not been programmed to deal with order mistakes. I asked to speak to a human. It ignored me. Again I asked for 鈥渁gent鈥�, 鈥渙perator鈥� or 鈥渉uman鈥�, in quick succession.
It failed to recognise my annoyance and the insults that followed and switch the call to a person (). Eventually, I hung up. Such systems are a misuse of speech recognition technology.
In some circumstances this technology can be good. I used speech recognition successfully a decade ago after breaking an elbow. However, I had to train that system for half an hour to understand my voice, and I took care to speak clearly and carefully into a good microphone in a quiet room. Human speech and accents differ widely enough for such training to be crucial if a computer is to translate speech into text successfully.
Customer helplines are a very different matter, because they must recognise what a caller says from the first word. That鈥檚 called universal speech recognition, and is much harder because it must accommodate the wide variation in accents and how voices sound.
The fact you are talking via a phone line rather than directly to the computer doesn鈥檛 help. The telephone network cuts off frequencies above 3400 hertz that are needed to distinguish between consonants such as 鈥渄鈥� and 鈥渢鈥�. on busy cellphone networks degrades quality still further.
Even when it comes to text-based systems, computers aren鈥檛 that good at understanding questions, as one quickly learns when trying to explain a problem to online chatbots. Humans remain much better at answering questions and resolving complaints. After my run-in with the automated system, I managed to reach a person the following Monday. He quickly understood the problem, sent me a code to download the proper software, and apologised on behalf of the computer.
(Image: Andrzej Krauze)
This article appeared in print under the headline 鈥淭erminate these bots鈥�
