Finding the right word
China now has technology that can transcribe Mandarin, Cantonese, and English sound bites into text and analyze their meanings. It’s a major breakthrough for business, though some are not pleased. Some dialect-speakers say its limited application threatens to marginalize their dialects. Miles Wen Haofu, founder and CEO of Fano Labs, is not worried. He’s fixing the problem.
The advantage of this new technology is that consumer comments on company hotlines can be recorded, and transcribed into text. Corporate research staff can review the messages and gauge consumer concerns efficiently. Yet, the system cannot transcribe many derivatives of Mandarin. Some people speak in dialects that are marked by strong accents, are not widely used and cannot be incorporated into the system in any practical way.
Wen’s AI-powered system has been upgraded to embrace Southwestern Mandarin, the branch of Mandarin spoken by 260 million people mostly in central and southwestern China. It also recognizes Chinese spoken by those who have spent time overseas and often inflect their speech with foreign languages. The system can identify English words thrown into Cantonese or Mandarin.
The system does voice-to-text translation using natural language processing technology, a branch of artificial intelligence. It teaches computers the natural way humans talk and turns the information into data for system training.
The system works much in the same way that children learn languages. Children hear people around them using words that eventually become familiar. Though not able to respond to the unfamiliar words at first, they absorb verbal cues, which form patterns in the brain, and language learning proceeds.
Speech recognition technology works in much the same way. Computers store huge files of sound samples linked to corresponding text. The computers are trained to identify the speech patterns and make predictions as to the most likely translation, Wen explained.
Fano Labs cannot match the advances made by tech giants like Baidu at the moment. While Baidu claims an accuracy rate of 97 percent for language transcription, Fano Labs still has a long way to go in improving its accuracy rate, which currently stands at 80 percent.
Though Fano Labs lags behind in accuracy, its ability to translate Southwestern Mandarin in China is unique, Wen said.
Basic recognition of sounds is not enough for the AI to carry out effective transcription. It needs to go further. Speech recognition systems have to be able to distinguish phonemes and homonyms to a much greater extent in Chinese than in English.
There are some spoken words in Chinese dialects that people can read but can’t write. These words are quite common but still beyond the reach of current technology.
“We cooperate with linguists to use the word with the closest meaning or sound to predict the caller’s meaning,” Wen said. The substitute words help the system place the words in context and achieve a clearer understanding of the caller’s intent.
Bound by a contract not to disclose the names of his corporate clients in Hong Kong, Wen said the Fano Labs system is already in use in the call centers of mobile network providers and property companies.
In Sichuan province — the main region that speaks the Southwestern Mandarin dialect — the company’s corporate clients include Sichuan Radio and TV Network, China Merchants Bank and China Telecom Global.
Fano Labs also includes staff training and evaluation as part of its service to back up voice-to-text translation. Taken as a whole, companies are able to analyze the transcriptions and determine the needs of callers, Wen said. The startup, a little more than 2 years old, has already attracted venture investments from Hong Kong and the mainland.
Last year, the company received investments worth millions of US dollars from Horizons Ventures, the venture capital firm which manages the private investment of Hong Kong tycoon Li Ka-shing. It was the first time it invested in a Hong Kong tech startup.
The startup today has more than 30 employees, and generated millions of Hong Kong dollars income last year. Despite all that, Wen told China Daily, Fano Labs is still in the red.
Even the most successful entrepreneurs have tales of woe about their early days. Wen, a confident straight-A graduate of the University of Hong Kong, was no exception.
Wen came from Jilin province in northeastern China. He was among the top 2 percent in the national college entrance exam in 2007, and was admitted to HKU on a full scholarship, studying in the Department of Electrical and Electronic Engineering. He graduated with the highest GPA in his major.
Wen’s tutor Victor Li On-kwok, chair of information engineering at HKU, suggested he study for a doctorate degree, to “know something better than anyone else on the planet”. Wen followed the advice and got his PhD at HKU.
When he started the business in 2015, Li and a schoolmate came on board. “AI may be a hot topic now but, five years ago, if someone said his or her PhD was in artificial intelligence, people would be skeptical,” said Wen.
People didn’t believe AI had much potential. In the first eight months into business, the entrepreneurs applied their technology to seven or eight projects. But none was sold.
By the end of May 2016, things got better and then much worse over a 24-hour period. On May 30, when in Chongqing for a Hong Kong Science Park event, Wen got a promise for his first contract. Vincent Lo Hong-shui, chairman of Shui On Land, was fascinated by his ideas.
Heading a company that builds shopping malls and residential properties in China, Lo showed interest when he heard Wen’s idea that customers can use AI to find places they were looking for in shopping malls — all they had to do was ask a robot, and a computer monitor would show the way.
Lo thought it was a great idea, and offered to collaborate with Wen. It was a major boost for Wen’s morale at a time when he was getting fatigued by the seemingly inevitable cold shoulder to his ideas wherever he went.
Wen was excited to share the good news with his partners, only to get shocked. The schoolmate, one of Wen’s partners, took the occasion to declare he wanted out and was leaving the partnership to take an offer from a big company that provides him with a solid income.
The partner had done all the programming up to that point. Wen’s tutor, Li, worked part-time and didn’t have time to write the code. Wen had to do it on his own — a massive project.
The day after breaking up with his partner, Wen drunk himself into stupor and woke up at 4 am the next morning, feeling driven.
From June 1 that year, Wen met people from Shui On Land to discuss big plans. He worked during the day and coded at night, sleeping less than two hours a day. That went on for six months.
Wen was nearly broke. He had only HK$1,000 in the bank and could barely keep up with his personal expenses.
Wen had spent almost all his money buying out his former partner, and had to borrow 30,000 yuan ($4,400) from his parents. “It was the first time I had asked for money from them since I came to Hong Kong,” Wen noted.
His partner’s hasty departure had put him on the brink of bankruptcy.
Having learnt the lesson from his experience with his ex-partner, Wen now makes it a point — always to have a contingency plan.
After six months’ work, the deal with Shui On Land fell through. Wen was able to carry on. Later, he took an order from China Light and Power, a major electric company in Hong Kong, while the Shui On project was still in development.
CLP wanted Wen to develop a chatbot for its call center, answering frequently asked questions. This time, everything came together and Wen’s reputation started growing. He had proven he was able to transcribe spoken Cantonese to text, even when it was mixed with English. Other companies with call centers in Hong Kong came knocking, looking for Fano Labs to build similar programs.
Wen saw endless potential in the dialect transcription market. While holding on to his edge in Cantonese, he wanted to explore a Chinese dialect spoken by many more people.
“Nearly 100 million people speak Cantonese around the world, while some 260 million speak Southwestern Mandarin in China, where I see a huge market,” said Wen.
Last March, Fano Labs set up a subsidiary in Chengdu, Sichuan province in southwestern China. The subsidiary company is developing an interactive device for the tourism bureau in Meishan, a city near Chengdu. “If a tourist wants to know the history of an ancient stone, he can ask, and our device will tell him the story,” Wen said.
He plans further exploration in other southern dialects, such as Southern Min spoken mostly in Fujian province and Hakka.
Li said this will hone their mixed language processing skills and bring him closer to his dream of dominating the Southeast-Asian market and, from there, the whole Asia-Pacific region. Wen hopes to grow Fano Labs into a unicorn company.
That goal is written in the company’s Chinese name You Guang Ke Ji — “you guang” means “having light”, and “ke ji” means technology. Wen said the aim is to use technology to illuminate blind spots in information technology.
As Wen looks forward to the future, he also looks back. He has hung on to the note he wrote in that memorable morning at 4 am, setting his own rules of engagement and driving him to keep on. As he processes more languages and dialects, he remembers the drive, that insistent voice, which keeps telling him to never give up on success.