To boost accuracy, Lingmo must regularly retrain its translation models—a task that takes up to 10 weeks using traditional machine learning techniques. How could it speed up the rate of improvement?
Lingmo switched to a cognitive approach, leveraging IBM Watson to enhance speech recognition and train models much more quickly, even with relatively small volumes of data.
9 languagescan be translated in near-real time, including dialects and local nuances
85% accuracyin real-world conditions helps users gain greater contextual awareness
50% fastermodel training boosts responsiveness to user feedback and enhances quality
Business challenge story
Learning to communicate
In a globalized economy, even small businesses are seeing the benefits of an international outlook—whether that means finding customers in new markets or sourcing products and components from overseas suppliers.
Meanwhile, for international enterprises, any technology that makes it easier to collaborate across countries, continents and cultures can potentially be a game-changer—providing a route into markets that competitors may not be able to access.
Lingmo International helps businesses take advantage of these opportunities by offering a range of next-generation translation products and services, including innovative mobile apps and state-of-the-art wearable translation devices.
Danny May, Lingmo’s CEO, explains why he founded the company: “A few years ago, I was in China and my passport was stolen. I tried to communicate with local police using a translation app on my phone, and… let’s just say it didn’t go well!
“It’s a common problem, especially for businesspeople who may be sent to work in another country at short notice, and don’t have time to learn much of the local language. There are a lot of translation apps that claim to help, but in real-world situations, they rarely deliver on their promises.”
May felt that he could do better: “Machine learning and artificial intelligence have a huge potential to solve these kinds of problems. For the first time, it’s possible to build accurate, nuanced, real-time translation services that don’t just give you the gist, but help you understand the full context of a conversation. By turning that vision into reality, Lingmo’s goal is to help people who need to make sense of a new culture quickly.”
Lingmo’s data scientists started by building a set of machine learning models that could be trained to deliver much more context-sensitive translations. The initial models were highly successful, and helped the company launch its first range of products and services.
However, there was one problem: machine learning is a very data-hungry process, and for the models to improve, Lingmo’s data scientists had to keep feeding them with large quantities of voice recordings. Sourcing this data became increasingly challenging, because many of the languages that Lingmo supports have a wide variety of dialects and local nuances, and native speakers of every dialect are not always easy to find.
“Continuous improvement is very important—in fact, it’s the whole purpose of machine learning,” says May. “When we get feedback from our users that our software isn’t picking up a certain dialect word or phrase, we retrain our models so that the problem won’t happen again.
“Unfortunately, the machine learning techniques we were using required so much data that it could take up to 10 weeks to crunch through a single training cycle. We wanted a more agile approach that could help us get better, faster.”
A meeting with IBM introduced the Lingmo team to IBM Watson, unlocking an AI-based approach that has transformed the way the company builds its translation services.
The Watson services help Lingmo split the problem up into different domains—using Watson Speech to Text for word recognition, Watson Language Translator for the actual translation, and Watson Text to Speech to deliver the result through an app or earpiece. Meanwhile, IBM Watson Natural Language Classifier helps the solution understand colloquialisms and dialect words.
“We looked at technology from a number of vendors who were able to achieve impressive results in lab conditions,” says May. “But in real-world tests, once background noise and other difficulties were introduced, they proved to be substantially less accurate than IBM Watson.”
He adds: “Watson enables us to train our models on text instead of requiring multiple voice samples of every word or phrase. That dramatically reduces the amount of data we need to feed in, and accelerates our training cycles by up to 50 percent.”
Using the Watson platform also enables Lingmo to leverage IBM’s leading-edge AI research, while safeguarding its own intellectual property.
“The Watson platform provides a smarter way to train our models on our data in the IBM Cloud—and at the end of the day, it still remains our data and our IP,” explains May. “In a world where data itself is a competitive advantage, it’s reassuring to work with a company like IBM that respects the importance of proprietary information and technology.”
The IBM Watson cognitive technology has been a key component in the development of Lingmo’s Translate One2One earpiece, which can translate between nine different languages (English, Mandarin Chinese, Japanese, Arabic, Spanish, Italian, Portuguese, German and French) with 85 percent accuracy. Content is translated and fed into the user’s ear in near-real time, delivering results within five seconds on average—and unlike most apps and other translation services, the earpiece does not rely on the user’s mobile phone to provide connectivity.
“With Watson, we’re often seeing 90 or 95 percent accuracy in practice, and we’re picking up dialect words that only locals know,” says May. “The feedback we’re getting from users is fantastic, because they can see that it’s a solution that doesn’t just work under lab conditions—it’s practical and useful in real-world situations.”
Whenever Lingmo wants to add a new dialect or improve the quality of its translations, the training cycles are twice as fast as before, helping the company integrate customer feedback into its products in a more nimble and responsive way.
“With Watson, we’re also able to take advantage of new data sources to make our translation services even more capable,” says May. “For example, we worked with IBM to teach our models how to watch movies, which gives us a whole new level of insight into cultural word usage and trends.”
Lingmo sees big opportunities for its products to transform a range of industries, and is seeing considerable interest from the travel and transportation sector in particular.
“Airlines need to help passengers from all over the world get to their destinations on time, and that can be a stressful experience if customers and staff can’t understand each other,” explains May. “Traditionally, airlines rely on recruiting people who can speak four or five different languages, but that’s not a scalable model, especially as margins get tighter.
“With our technology, we can effectively give every member of staff the ability to understand nine languages, which lowers the bar for recruitment while improving the level of customer service. Whether a passenger is panicking that they might miss their flight or just wants to order a drink in the lounge, the ability to communicate effectively makes a big difference—and could be the key to retaining them as a loyal customer.”
Above all, IBM Watson technology is helping Lingmo fulfil its mission of helping people communicate and feel comfortable, no matter how far they are from home.
May comments: “We have a Spanish professional soccer player who recently came to play for a team here in Sydney. When he arrived with his family, his wife couldn’t speak any English, so we lent them a prototype of our next product. Now she feels much more comfortable about leaving the house, taking their baby out, and generally making a life for her family in a new community.”
He concludes: “We’re glad to be working with IBM to develop cognitive translation solutions that make a real difference to people’s lives. It’s more than just removing language barriers – it’s enabling and empowering more people to communicate with each other and open the door to new possibilities.”
Lingmo International is a leader in translation technology, focusing on speech recognition, accuracy, and perfecting the interpretation of dialects and nuances. Headquartered in Sydney, Australia, the company offers consumers and organizations smart translation solutions across voice and text platforms, and operates across Oceania, Asia and North America.