If Alexa, Amazon’s voice assistant, is capable of listening and speaking the credit is largely his. And Prem Natarajan, an Indian-born scientist on loan to the business, has one goal: democratize artificial intelligence. Starting with a very simple idea: to have the users themselves teach her new things.
Alexa is one of Amazon’s most successful services: the digital assistant is considered by analysts later compared to the competition from Siri and Google, and Amazon considers it one of the most important tools for the relationship with customers of the company founded by Jeff Bezos. Italian Tech exclusively met Prem Natarajan, vice president responsible for Alexa AI, during the annual re: Mars conference – dedicated to the topics of Automation, Robotics and Space – taking place in Las Vegas. “One reason for the results we have achieved – says Prem Natarajan – is linked to way in which we think inside Amazon: “Working backward”, working backward “. We start from what we want to give to customers and go back, building the technologies necessary to get there “.
In the case of Alexa has been a long way. Amazon bought it in 2013 from a Polish company (at the time it was called “Ivona”) and which it launched on the market in 2014 with the first generation of the Echo and Echo Dot smart speakers. Since then, Alexa has grown a lot (“If she were a person, you wouldn’t recognize her,” Prem Natarajan jokes) and more it is used in a transversal way on many products of the company, but acquired increasingly sophisticated functions. Like the last one announced during re: Mars.
What is asked of Alexa: here are the questions of the Italians
It will take you a minute of audio to “learn” to speak with anyone’s voice (previously similar systems required up to an hour of prepared audio test). On the re: Mars stage another executive responsible for Alexa development, Rohit Prasad, showed the example of a little girl who asks Alexa to read her the story goodnight in the voice of the deceased grandmother. The goal, Prasad said, is to “make memories last” after “so many have lost someone they love” during the pandemic. On a lighter note, Alexa also prepares to go to space as part of Orion, a NASA demonstrator of the Callisto I mission. The project? It’s titled “Alexa, take me to the Moon”.
But how does Alexa work? How do you listen, understand and respond to our requests?
“The journey has been long. Speech recognition was initially considered an absolute challenge – Prem Natarajan says – but not only have we solved it, but we have also worked on a natural way to “listen” to people’s voices, because people don’t want to keep a button pressed as if they were talking on the radio “.
World Voice Day
The Day of the Voice between podcasts, chats and smart assistants
by Simone Cosimi
The idea of Jeff Bezos to which to arrive was the computer of Star Trek, an invisible entity present on the bridge of the Enterprise that Captain Kirk summoned with a “Computer, how far is Planet X?” and received a quick and focused response. Alexa was basically born with this goal. “Starting from this – says Prem Natarajan – we have worked hard to eliminate all friction points and defects for customers: we work with models that have all languages in parallelthere are 17 in more than 40 countries, and this approach works much better than dedicating ourselves to a single language. “Indian, Italian and French are being developed hand in hand with English to give Alexa even richer and richer skills.
Over time, let there be progress in computing power that in the machine learning algorithms has allowed to compact the needs of Alexa to give an answer. And the modular approach, based on “skills”, ie on the specific skills that Alexa receives both from Amazon itself and from third parties, allows for timely information on very different topics. As soon as it is turned on, Alexa is already able to answer many questions on the most disparate topics, taking advantage of the former online database of WolframAlpha, iMDB, AccuWeather, Yelp, Wikipedia and many other sources of structured information. But this is only the beginning.
After Alexa and Google Assistant comes Sonos’ voice assistant
by Bruno Ruffilli
“We use artificial intelligence to improve artificial intelligence, ”says Prem Natarajan. This happens both in the voice recognition part and in the dialogue modeling part: Prem Natarajan has worked to generate billions of synthetic conversations, built by automatic systems and supplied to neural networks to make Alexa “learn” how to hold a more refined conversation, that doesn’t sound “robotics”.
“There are three central aspects – says Prem Natarajan – which should not be misunderstood: for Alexa and AI in general, I am talking about the era of the “self” of AI. This means “self awareness, self learning and self reasoning”. Attention, I’m not talking about living AI or other naiveties that are read in the newspapers in this period: they will be research problems for our children, certainly not for this generation “. The reference is to the story of the engineer Blake Lemoine from Google who declared that the AI he works on has an “emotional sensitivity” and he is basically a person.
Alexa, where do I go on vacation?
Prem Natarajan’s approach is completely different and much more pragmatic: Alexa is trained to learn new things by herself and to activate herself in such a way as to be able to predict and respond better to even implicit requests from her “masters”. Alexa is learning better to understand how the person to whom she has to respond speaks, to understand sophisticated concepts such as “Alexa, regulate the temperature in a pleasant way” and understand what “pleasant” means”For that single person it requires some automatic passages to Alexa but next time she knows.
“In the future, Alexa will become more and more able to move independently in environments such as the home, learning not only how to do it but also what to do: check the windows or figure out if the lights are on or off at certain times” , says Prem Natarajan. This expansion of Alexa, from space to Astro, the little robot that has been on the market for six months a little at a time by invitation in the US and which represents what Amazon believes will be the near future of the smart home, is linked to the Alexa Everywere project, coordinated by Aaron Rubenson. The goal is to make Alexa Amazon’s “unique voice”.
The voice service “Inps News” for Alexa and Google Assistant starts
“More and more things can be done in the future – Prem Natarajan – how to ask Alexa “Check if the door is closed” without giving her any other information, perhaps while one is traveling, and she can understand which door we are talking about and how to check if it is closed. A key to this evolution is the ability to train, or as Prem Natarajan says “educate” Alexa.
Amazon’s goal for Alexa is “generalizable intelligence”, that is, the ability to adapt to user environments and learn new concepts with few external inputs. But it is not to be confused with the omniscient, all-encompassing and super-artificial general intelligence, which are instead trying to develop those of Alphabet’s OpenAI and DeepMind, co-founded by Elon Musk.
The central point, according to the scientist, is to work on the ability of Alexa (and other AI) to be aware of where it is (thanks to sensors and other types of data), to learn by itself (thanks to algorithms that allow to acquire and process new information depending on the context) and to have a conversation “reasoning” with his parents interlocutors and not simply by answering direct and univocal questions.
The latter is the social aspect it takes to make Alexa one instrument useful but that should not deceive us: “Alexa is not alive, sentient AI is not something that concerns us, at least for a few more generations. Instead, I think the value and comfort that Alexa brings to everyday life is really interesting ”.