How Amazon’s Alexa was taught to understand the Scottish accent
SPECIAL REPORT
HE’S the man who taught the Scottish accent to Alexa, the voice of Amazon’s Echo digital assistant, so that she can respond if you ask her to turn on the big light, “gie’s a bit of Beyonce”, order pizza crunch from the nearest chippie, or find out if it is dreich outside.
His task involved many darkened rooms, subterfuge on an industrial scale in co-opting scores of unknowing participants, dozens of hard disks air-freighted to Silicon Valley and hundreds of thousands of pounds of expenditure. It was, in the argot Alexa is now fully au fait with – a pure, dead brilliant operation.
Back in the summer of 2015, Richard Thynne was looking for a well-paid temporary job. He has a post-graduate degree and was thinking about teaching but needed something to do over the summer before he went travelling. He approached the graduate recruitment company Freshminds in London.
“I was asked if I had three months on my hands, which I did,” he recalled, and so the cloak-and-dagger adventure began. “The person who interviewed me only knew it something to do with the technological aspects of data capture, not who the client was. That was only known to one person in the company.”
The next stage was a series of phone interrogations down the line from California, one from the person leading the project, Nehal Meshal, described by Amazon as a programme manager.
It was only when he was selected and turned up again at Freshminds to join seven others, the final team, that he was told the client was Amazon – not what they had been selected to do.
A week later he was in a boardroom in the company’s massive 200,000plus square feet glass and steel London headquarters in Holborn Viaduct. There he was told in detail about the Alexa project, but only on condition of strictest confidence. He and his colleagues were to go round the country sampling regional accents so the digital assistant they planned to launch in the UK could understand the many accents of the British isles and respond properly.
“No-one in Amazon in London knew what we were doing or why we were there. We operated from a blacked-out room. The US guys had hired a mar- ket research company here, Appen, but only one person in that company knew what was going on and he had been flown in from Australia, presumably so he couldn’t blab over beers to associates in London.”
It took a week to be trained. “It was quite a complicated affair, with loads of equipment, 20 mics at different levels and heights in different parts of the room, with all the co-ordinates then fed into the computer to provide a 3D map of the room, along with photographs.” All of the data was then fired over to Silicon Valley.
It took a day to set up the equipment, and volunteers, surrounded by a battery of microphones and speakers, were paid £40 to take part.
Voice recognition software works by analysing the airwaves you make when you speak, translating that into digital data. Samples of the waves are taken at frequent intervals – hundredths or thousandths of a second. The higher the sampling and precision rates, the higher the quality. These samples of words are then compared with the same words already stored in digital templates and matched, modified, together with other elements of technological wizardry known only to the 1,000 or so developers who worked on Alexa.
Thynne and his colleagues were paid £120 a day, plus £20 food expenses, and put up in good hotels. “This was no cheapskate, exploitative venture,” he said.
The first stop for Thynne and the team was Manchester. The bloke from Oz, Richard Ingram, was the trailblazer going on ahead to each of the designated cities. There he would book two flats or houses through Airbnb – the group had now been split into two teams of four – and line up participants for the study, who were not told why they were taking part.
“We had a spiel and if people asked what it was all about, I told them we were measuring how sound rebounded in rooms. It was a complete red herring,” said Thynne.
Those taking part had been asked to come with the titles of five books, five films and five favourite pieces of music in their heads. The sessions were in three parts. “They had to read a series of prompts, press the button on the tablet when they were ready and speak them out. There were 40 or so, like ‘Find me the weather in Manchester’.”
IN the second phase the prompt left a blank at the end, such as “Find me a book by ...”, with the person adding in their own favourite. Finally, each participant asked their own questions. To simulate a real household environment Thynne and the others would occasionally switch on a TV or a washing machine as the person was speaking. There were between 200 and 300 prompts and responses in each session by each one of the more than 1,000 people taking part, all of it recorded and shipped off to the States. “A mind-boggling amount of data,” said Thynne.
As there was media talk about Alexa’s launch in America, Thynne said: “Some people guessed what we were up to but the overwhelming number didn’t. We had a large box of hard disks and every two days we’d trot to the post office to send them to London,” he added, “and from there to California.”
After Manchester it was Leeds for two weeks, followed by the fortnight in Glasgow, the same in Birmingham then the last stint back in London. While in those five main cities, the team also recorded subjects who had Irish accents, Polish accents and “others for whom English wasn’t the native language, a whole range”.
Clearly Amazon was pleased with the result. Alexa launched in the UK last September, with an English accent, and amid a blizzard of TV and newspaper advertising. Thynne would like to make his own judgment, but despite all his hard work Amazon didn’t send him his own digital helpmate.