“Namaste, main Alexa hoon. Naam toh suna hi hoga…,” an Amazon Echo device proclaimed as the e-commerce giant announced the launch of its voice assistant in Hindi and English.
For Rohit Prasad, vice-president and head scientist of Alexa AI, it was the culmination of a long journey, where Alexa users in India themselves helped by volunteering to teach the system phrases using the Cleo skill on the platform. It also helped that Alexa had been live in Indian English for over two years. “This meant a lot of the content names was already identified,” he told indianexpress.com.
Prasad, who leads Alexa research and development in Artificial Intelligence technologies and holds several patents in this domain, said the idea was to see how they could model it with minimum human supervision. “What if we could just train a deep learning-based model that learns all of these phenomena automatically, right?” The deep learning models were also able to learn pronunciation variations. This also meant it was easier to address Hinglish queries, Prasad said. However, Alexa did have to learn Devanagari script from scratch since this was new for it.
Users will be able to use Alexa in Hindi by just changing the language in the backend. Soon, Alexa will be able to switch between English and Hindi seamlessly, Prasad said, adding that this will help multilingual users and households.
Chipping in, Puneesh Kumar, Country Manager for Alexa Experiences and Devices, Amazon India, said they built the Hindi language product in record time because users were already interacting with a mix of English and Hindi. “This is a perfect example where a lot of things we’ve done for the first time. And the benefit we’re getting is not just within India, but will also help with other launches and languages in the future.”
Prasad said it was important not to miss the out part of the virtual assistant also, as something pronounced by Alexa wrongly could have a negative impact with users. “That is another daunting challenge where you have work cut out.”
Having done 15 languages before Hindi, Amazon also leaned on multilingual modelling. “Deep learning makes that quite natural, because you can train a big, massive deep learning-based model that trained on all languages, and then adapt only the last layer to the target language,” Prasad said, adding how that helps get better in new languages faster.
Kumar reminded that while we look at Hindi from the Alexa perspective, it was also about Amazon. “We are always going beyond the next hundred million customers and always looking to go beyond English. What the customers are telling us, kind of guides the priority for us. We are very customer-obsessed and truly listening to what customers are saying,” he said, adding how the learning from how customers are interacting on the platform has given them clarity on what should be the “most important piece right now”.
Explaining how user behaviour is not the same across the devices they use voice in, Kumar clarified: “We don’t want to teach people how to talk to devices, we need to learn the diverse ways in which people talk naturally. Alexa being in the cloud, should be then aware of the context.”