Will Vietnamese companies be sidelined in the AI race?

The debate between Jack Ma and Elon Musk at the World Artificial Intelligence forum earlier this month set Vietnamese social media abuzz. In a country with a burgeoning tech-savvy population, it’s understandable that AI is being discussed everywhere – from government officials’ speeches that cover their usual Industry 4.0 rhetoric to companies and corporates proclaiming that their products and services are powered by AI.

Photo credit : joruba / 123RF

The entire Vietnamese tech scene is still evolving, and AI development in the country is still at a very early stage. Local companies and startups are racing to keep up with their foreign counterparts.

“The opportunity is there if we can understand our customers better and harness the data that we have and Google doesn’t,” said Do Van Hai, a speech processing expert at the Viettel Cyberspace Center. He has been leading efforts within Viettel, the country’s largest telecom provider, to apply AI in building speech applications.

Natural language processing for the Vietnamese language

Formally established in 2015, Viettel has been tasked with research and development for applications in big data, AI, data mining, and deep packet inspection. For now, the center’s AI section particularly focuses on speech processing to improve Viettel customers’ experiences and harness the potential of telecom data owned by the conglomerate.

Viettel’s voice recognition service is considered tip-top in the country. When it comes to processing the Vietnamese language, Hai said, Viettel’s algorithms are more accurate than Google’s. Originally, Hai’s team was set up to analyze data from Viettel’s customer service center, which often receives 500,000 phone calls in a day, as about two-thirds of Vietnam’s population subscribes to Viettel’s services.

In the past, content of customers’ phone calls to Viettel were checked by the company’s top leaders randomly to develop a better understanding of customers’ complaints. Progress made in speech recognition and data mining over the last few years has allowed Viettel to monitor half a million phone calls from customers – transcribing the conversations if needed – to detect snags.

Globally, virtual assistants have not been able to fully replace call agents. His team, however, saw the potential in transferring voice recognition services to other platforms. Viettel’s applications have been used by several online newspapers for text-to-speech functions, and they were adopted by state agencies to automatically transcribe meeting notes. In the future, other applications could also include virtual announcers and automatic movie voice-overs in Vietnamese.

“Obviously, we still have a lot of work to do to improve the level of accuracy,” Hai said. “But when it comes to the Vietnamese language, we do have some competitive advantages.” He was referring to access to data that is not available in the public domain, as well as preferred access to contracts offered by the state, which will only use domestically developed AI solutions due to national security concerns.

Take your pick

Smaller players have also made fresh developments in this arena. Tuan Trinh founded NextSmarty in 2016. The startup applies AI and deep-learning algorithms to build recommendation services, targeting medium enterprises that do not have the resources and data to develop their own algorithms like those of Amazon or Netflix.

In 2017, Tuan co-authored a paper detailing how his solution generates real-time recommendations even if a user does not create a profile, or when a platform doesn’t have enough data that covers the user’s past transactions. Simply put, Tuan developed a method to capitalize on the order of a user’s clicks.

He recalled that the idea first materialized when he was working as a deep-learning engineer in Silicon Valley.

“You can always find something on Amazon, but what about a much lesser-known website? You can easily leave after two or three clicks. This is the pain point for many websites that do not have enough data on user interactions in the past and with far fewer loyal customers,” Tuan said.

NextSmarty’s first client was a foreign luxury jewelry ecommerce website, which according to Tuan was struggling to increase sales because conventional AI-powered recommendation approaches did not work. There simply wasn’t enough data that covered user interactions for the AI system to interpret because “most users don’t buy jewelry very often.”

Building digital identities

Still a long road ahead