Indian linguistic diversity challenged Alexa to be better: Amazon executive

BENGALURU: India’s linguistic diversity made it one of the most challenging markets for Alexa, but pushed Amazon’s voice assistant to also become a fundamentally better product, a senior executive at the US-based technology giant told ET.

India has an incredible diversity of languages, vocabulary, dialects and pronunciations. Besides, Indians tend to pose mixed language queries to Alexa – a behaviour that is unique to India.

“Developing Alexa for India was the most challenging experience for language processing, but also the most rewarding experience,” said Prem Natarajan, Vice President, Alexa AI & Head - Natural Understanding. “Awesome thing is that when you are challenging problems for a diverse country like India, it turns out you are developing techniques that benefit others across the world.”

India was the fourth country after the US, the UK and Germany in which the voice assistant was launched, back in October 2017.

Amazon wants artificial intelligence-driven Alexa to be an extension of its product offerings of ecommerce, cloud, music and entertainment.

“A language can exhibit variations in dialects, vocabulary, accent, pronunciation. In most parts of the world, you see broad clusters of these. In India, language is individual. It is far more granular. This means, the learning challenge in India is much greater,” Natarajan said.

In other parts of the world, Alexa looks at natural language understanding as single language modelling or processing, but in India it thinks of it as multi-lingual modelling or processing for every component of the system.

Every time Alexa is asked a question, it converts the question into text to search for an answer. Once Alexa gets the answer, it converts the text back into speech.

“The underlying technology is neural text to speech, but instead of using a single language model during generation we have to use multiple language models and shift back and forth between different language models,” Natarajan said.

Another challenge that Alexa specifically faces in India is a lack of standard English spelling for regional words. There are multiple spelling and pronunciations for names, movies, songs and places. To solve for this, Alexa had to develop a deeper directory of phonetic search for India.

“In the West, there is a lot of standardisation. Search problem in India becomes a science problem because I have to come up with ways for search techniques to be robust with variation. In India, you can’t have a complete vocabulary,” he said.