Welcome to the age of the ‘superhuman’ chatbot
Computers use up to 1,000 separate voice markers to make judgements about us – including how much you’ve drunk, whether you show signs of autism and even guess your height
Chatbots already have ‘superhuman’ qualities that can guess your height, recognise raised heart rate or alcohol levels and even early signs of autism that would not be picked up by the human ear, a leading expert in AI and machine learning says.
Audio computer systems are also able to recognise more and more about you – from your personality, your gender, to your health state, says Professor Björn Schuller, an expert in machine learning at Imperial College.
“Computers are now superhuman in how they respond to the human voice, for example recognising if someone is alcohol intoxicated. Computers are now better than humans are at listening to the human voice at subtle levels of blood alcohol concentration.”
“Just from listening to your voice, computers can estimate the length of your vocal tract and how that correlates with your body height.
“It even goes as far as early recognition of Rett syndrome or autism at an age where doctors usually would not have assessed them, from pre-scanning of the voice,” he said.
“Computers could also rank me as a native German speaker, somewhat accurately estimate my age and even my educational background.”
“It is very rich audio analysis. Anything that a human can do, a computer should be able to infer from your voice as well, in practice, deep neural networks are also using up to 1000 acoustic features alongside ‘end to end’ learning from data”, he said.
Sophisticated new ‘virtual agents’ can also recognise emotion, frustration and anger in human voices alongside other ‘bio-signals’ that form part of a new branch of research called ‘artificial emotional intelligence’, according to Professor Schuller, the chair of embedded intelligence for Healthcare and Wellbeing at the University of Augsburg.
Professor Schuller, who is also the founder of a firm called audEERING that develops computer-aided listening techniques, said the possibility now existed for ‘genuine’ or “well coded emotion or personality” to learn to respond to the human voice.
“Even in the short run, these systems are now able to robustly recognise basic and complex emotions and personality traits such as openness, conscientiousness, extroversion, agreeableness and neuroticism.”
“We do what we call ‘artificial emotional intelligence’. This includes not only recognising emotion from the sound of the voice and the spoken words, but giving machine algorithms emotional-like behaviour and emotional-like learning abilities to exploit similar effects to those that emotion has had in human evolution.”
“These emotions currently trigger automatic dialogue agents, so the very easy thing is to try to change the dialogue to calm the customer down and also to adapt the speech recogniser,” he said.
Techniques include lowering the voice or using a more soothing tone are being adopted, he said, while technology that recognises anger is already being provided to global technology players. The result means scientists may soon have to consider the ethical issue of how to differentiate between a human voice and a synthesised one, says Professor Schuller.
The issue will become important once virtual agents can bridge the ‘uncanny valley’ – a phenomenon where humans recount feeling increasingly positive towards human-like robots, until they are very close to human-like behaviour, but not perfect, yet, where acceptance breaks away.
“Once we are over this gap, the ‘uncanny valley’, it’s likely the voice will be so human-like that you have to make people aware of it,” Professor Schuller says.
The impacts of more sophisticated ‘virtual agents’ could be widespread on the world’s $310 billion a year call centre industry – and may even become a trigger point for national economies as they consider whether to introduce Universal Basic Income.
The sector is predicted to be one of the first to face widespread automation and is among a range of customer services occupations predicted to have a 91% likelihood of being automated in the near future, according to a recent joint study by Oxford University and Deloitte.
Recently, the International Monetary Fund found that Universal Basic Income could be effective in reducing inequality by protecting people from technological change.
“If we cross this ‘uncanny’ bridge, then virtual agents can become superhuman,” Professor Schuller said. (A machine that exceeds qualities found in humans)
The advantage would be that computers could always “remain nice” and their extended focus would allow them to pick up nuances that human operators might miss, he said.
“At some point they will be able to influence you or talk you into buying something, in a better way than even a good psychologist could,” he said.
The idea poses the question raised by Dr Yuval Noah Harari, a Professor at the Hebrew University of Jerusalem, this year in his best-selling book, Homo Deus: If computers can replicate organic algorithms – could they some day become better at manipulating human emotions than we are ourselves?
It’s reckless to forget people when we talk about AI. Here’s how to put humans at the heart of tech.