02/23/19 KISSING KANJI GOODBYE? (int'l edition)

Return to main story

KISSING KANJI GOODBYE? (int'l edition)
Asia's many thousands of characters are a huge headache
In Japan's parliament, stenographers sit in the middle of the house, recording every word on paper. After the session, they go back to their offices and punch it all into PCs. Not exactly a model of efficiency, but there's little choice: The Japanese language is so complex that no one can type it fast enough to keep up with speech, even at the sleepy pace of the Diet. China has it worse: There's no phonetic alphabet as there is for Japanese, and, besides, most citizens can't type anyway.
The fact is, what the West takes for granted--the ability to put language in to digital form--is a vexing problem for cultures whose writing is based on thousands of ideograms, the symbols used in Chinese, Japanese, and other Asian languages. Chinese uses between 4,000 and 6,000 characters, and each character can take a dozen strokes of the pen. Japanese uses somewhat fewer of these kanji--around 2,000--but it relies on three other alphabets in parallel: the Roman and two phonetic alphabets known as kana.
The implications go far beyond recording minutes of a meeting: The complexities of kanji-based languages prevent countries such as China and Japan from fully taking advantage of the Digital Age. And now more than ever, that's something they can't afford.
The answer may lie in speech recognition, which bypasses keyboards altogether. The technology is starting to gel. In the past year, products have quietly begun to appear that promise to pierce the language barrier. ''We think this will significantly help the growth of the Chinese software industry,'' says George Wang, director of the IBM China Research Laboratory.
Last year, for example, NEC Corp. introduced a product that, for $80 (with headset and microphone) lets users dictate and manage E-mail by voice. The product generating the most buzz, though, is IBM's $99 ViaVoice, a continuous-speech program. It went on sale last September in China bundled with high-end PCs from Great Wall, Legend, and Tontru. A Japanese version was introduced in December and ships with IBM's Aptiva line of PCs. It converts 200 to 350 syllables a minute, about the normal speaking speed.
Still, the Japanese and Chinese languages present special challenges for speech technologists. Both are riddled with homonyms--words that sound the same but have different meanings. And speech-recognition programs must catch the different tones and regional dialects of Chinese. Moreover, both languages string words together without gaps, making parsing even harder.
Engineers hope to overcome the obstacles by sheer force: throwing ever-faster chips and vast language samples at their talking machines. At IBM Japan's Tokyo Research Laboratory, volunteers were asked to read a total of 35,000 sentences to help establish how they cut off words. The IBM team then copied 5 million sentences from newspapers into a computer so it could learn what normally follows any given word and had 1,000 people read from newspapers so their voice patterns could be recorded.
MAIL CALL. The result is impressive, but it doesn't go all the way. In China, ViaVoice still requires speakers to abandon regional dialects and to speak in clear and grammatical sentences: The ''ums'' and ''ahs'' of everyday speech turn up as unwanted characters on the screen. Still, IBM's Wang scored 87% for accuracy on his first dictation. After an hour session reading to the system, this rose to 95%.
Many of the first really useful products will be focused on specific tasks. NEC, for example, has developed a voice chip that recognizes 1,000 Japanese ''words'' averaging five syllables. Pioneer Electronic Corp. uses the chip in a voice-operated car-navigation system.
At the research and development level, NEC is creating a voice-based system to help the blind navigate a PC--and to have the computer read aloud mail and other documents. And Kyoto's Advanced Telecommunications Research consortium (ATR) is developing speech-recognition techniques for automatic translation systems--a Holy Grail for legions of monolingual Japanese on overseas trips. ATR scientists have developed a program that translates a spoken inquiry--even if poorly phrased---about hotel room availability. ''We're not so concerned about a broad vocabulary,'' says engineer Yoshinori Sagisaka. ''Our main target is spontaneous speech.''
One day, that might provide some relief for those parliamentary stenographers--not that it will help make sense of the politicians.
By Sebastian Moffett in Tokyo

Return to main story