Moving to a new country can be hard. You don’t know the language. Cultural differences create conversational landmines. And you just can’t be sure that everyone will like you. As it turns out, that as true for people as it is for Amazon’s Alexa voice assistant, which officially sets up residence in France today.
Amazon announced its expansion into France last week. The first wave of Echo devices ships today. But the task of prepping Alexa for its debut there began much longer ago. And how it learned not just a new language but an entirely new set of perspectives and priorities can tell you a lot about how Amazon will translate its homegrown voice assistant into a global success.
We Come From France
Alexa already spoke English, German, and Japanese before the France launch. But Google Home’s Google Assistant spoke all of those as well, along with French and Italian. It’s a global linguistic arms race. Winning hinges on getting it perfect, fastest.
When you think about what it takes to launch Alexa in France, start with the basics. There’s the language, obviously. But unpack that: French is complex, both linguistically and societally. It has formal and informal address. It demands of its speakers euphony, harmonious and seamless transitions between words to maintain an almost musical cadence. And as you might expect from a country with nearly 70 million inhabitants, a multitude of regional accents inform pronunciations.
Teaching Alexa how to speak French, in other words, is about as far from basic as it gets. You can feed the machine learning models all the French words in the world (literally), but teaching it how to use them requires a human touch. Like any good problem that machine learning solves, it starts with data.
“Amazon and actually everyone else of the big players will have a French language decoder available that does a pretty good job of recognizing speech, a particular language. That’s the foundation,” says Alex Rudnicky, a speech recognition expert at Carnegie Mellon University. “Once you have that, you have to figure out how people might talk to something. Then it gets a little trickier.”
‘We have different ways of asking the same question. We have different ways of saying the same thing.’
Nicolas Maynard, Amazon France
For Amazon, that meant introducing Alexa to workers in its five French fulfillment centers, who interacted with the burgeoning voice assistant in Boigny-sur-Bionne and beyond. From there, it broadened out to French Amazon customers who were granted early access to a voice assistant in their mother tongue. The conversations were unscripted, to let Amazon’s AI understand not just how it could best answer questions, but what kinds of questions were more common. Cooking, perhaps not surprisingly, came up frequently. But early Alexa users in France also sought out information about television a surprising amount, prompting Amazon to partner up with a popular French magazine to have a TV-related skill ready at launch.
More fundamentally, all of those interactions helped the algorithms behind Alexa to identify the many flavors of French-speakers. “In terms of accents, it’s very hard to say how many different ones,” says Rohit Prasad, head scientist of Alexa AI. And that means more than just the stark differences between what you’ll hear in the North of France versus the Southeast. “You can have people who have moved from the Middle East or India.”
Rather than slog through the process of modeling every single accent and variable, Prasad says, Amazon has set Alexa’s parameters broadly enough to understand voices from Le Mans to Marseilles.
Alexa also must also then contend with English words that the French language has absorbed; “weekend” is in common usage there, but with a French accent. And that’s before you even get to musical artists with English-language names that French pronunciation transform into an almost entirely new word, like Radiohead, or Earth, Wind, and Fire. (No, really.) For that, trial and error with beta testers has gone a long way.
After all that, you wind up with an Alexa that hopefully understands what people say when they speak French to it. But understanding isn’t anywhere close to conversing.
Duration and Time
Every language has its quirks. French just happens to have especially problematic ones for a voice assistant. “You cannot just take everything that is in the US, do a translation, and expect it to work in France,” says Nicolas Maynard, country manager for Alexa France.
Take the word “you.” In English, pretty straightforward. But in French and other Romance languages, it comes in formal and informal varieties. In France, Alexa always uses formal address; it’s there to help people, not befriend them, and better not to risk offending customers. But plenty of people engage Alexa informally, a distinction that has repercussions for comprehension.
“We have different ways of asking the same question. We have different ways of saying the same thing,” with the pronunciation varying depending on if you’re being polite or not, says Maynard. Alexa needs to understand both.
Amazon has set Alexa’s parameters broadly enough to understand voices from Le Mans to Marseilles.
It gets more complicated still. English uses different words to indicate time versus duration: o’clock and hours. French carries no such distinction, and you can imagine how that plays out when setting a timer, or an alarm. The things you take for granted in one country become potential potholes in another. And that’s just the grammatical whirligigs Amazon’s team has covered in advance.
“The tricky bits are figuring out exactly how people express a concept. The other bit is what is it that people do, how do they go about performing tasks in different cultures,” says CMU’s Rudnicky. “You have to have a sense of how people think of what it is that they’re doing before you can put something together.”
And not just what they think, but what they care about. Obviously French Alexa needs to know soccer, but it also boned up on rugby before its debut. Not to mention another area of particular interest to the French audience.
“French people do love to know when are the holidays,” says Maynard. “But the holidays are split in different zones in France. Depending on what city you’re in, the holidays are not at the same time. So we had to teach Alexa the different holidays in the different zones.”
What else? Generations of French school children have grown up memorizing the poems of Jean de La Fontaine and others; Alexa knows them, too. Alexa France has its own favorite movie (Stephen Spielberg’s A.I.), a favorite in the World Cup (France, obviously), a favorite singer, favorite book, all carefully chosen to reflect the broadest possible appeal. That extends to Alexa’s voice in France, selected as much for its sparkle of personality as its totally neutral accent.
Helping Alexa’s entry into France will be 200 localized skills that scratch the itches of early testers. Amazon has signed on the expected suspects; Air France, the leading newspapers Le Figaro and Le Parisien, popular radio stations, and other assorted content providers, all tailored for the market. To each culture its own voice-assistant needs. “In certain countries, certain skills are more important than others. In Japan and the UK, transit skills are important. In the US, maybe less,” says Toni Reid, Amazon’s vice president of Alexa experience.
So yes, Amazon has put in the work to make Alexa as French as possible. The question now is how quickly it can adjust for all the things it didn’t account for. “With any complex computer system, there’s always something you didn’t think of, and of course people pick up on it,” says Rudnicky.
Having been through these launches before, in Germany and elsewhere, Prasad has confidence in Alexa’s ability to adapt. “Once we launch, just like what happened in the US and other countries, you get more advanced data, more contemporary data, the models start to improve quite quickly,” he says.
And if it doesn’t? C’est la vie! At least it’ll have La Fontaine.