GC, Here's a current article which is interesting too!!
INSIDE TRACK: A system that is as good as its word:
TECHNOLOGY SPEECH RECOGNITION: Talking to a computer could become commonplace, writes Geoff Nairn
82% match; Financial Times ; 08-Mar-1999 02:21:36 am ;
1143 words
Speech recognition has come a long way in a short time. A technology once plagued by poor performance and high costs has improved dramatically in recent years, and talking to a computer, either directly or over the phone, could become commonplace in the approaching digital age.
For many, speech recognition will be forever associated with Hal 9000, the talking computer in Stanley Kubrick's classic film 2001: A Space Odyssey. But building a Hal-like computer that can understand the words and context of any conversation - "unconstrained speech" - is
a Herculean task.
"Just look at how much effort was required for a computer to beat a grandmaster at chess," says Stuart Patterson, president and chief executive of SpeechWorks, a US start-up developing speech recognition for customer service applications. "Researchers have been working on
unconstrained speech recognition for many years, but a computer similar to Hal is still decades away."Speech recognition technology works best when the vocabulary is limited and the dialogue follows a script. A typical application is phone banking where the dialogue is fairly predictable. Nevertheless, the system must understand a wide variety of accents and differentiate between "Dollars 40" and "Dollars 14", a problem exacerbated by the poorer quality of telephone speech.
SpeechWorks' technology draws on research licensed from the Massachusetts Institute of Technology, where speech recognition has enjoyed intense research. But only recently has the technology matured sufficiently to attract commercial interest outside the laboratory. An important catalyst is the power of today's PCs and their built-in sound support, which has encouraged the development of PC-based speech software priced to attract a consumer market. "There has been a tremendous amount of progress in speech and language technology development in a relatively short time," says Gaston Bastiaens, president and chief executive of Belgium-based Lernout & Hauspie.
L&H is one of several companies offering PC speech software and it has attracted funding from Microsoft, which plans to incorporate speech in its products, although it is cautious over the exact date. Other contenders in this fast-growing market include Dragon Systems, a US pioneer, and mainstream IT vendors such as Philips and IBM. PC speech programs have claimed accuracies of 95 per cent and recognise continuous speech, so eliminating the need to pause between words - a big drawback of early systems. Speech recognition is not easy. The sounds must be converted into a digital signal and then processed using a technique called spectral analysis. Finally, the phonemes - basic building blocks of speech - must be identified and grouped as meaningful words.
Recognising continuous speech is even harder because it is difficult to identify the start and finish of each word. Also, the pronunciation of each phoneme can be modified by those that come immediately before and after. Some languages suffer more than others from this effect, so companies have to make significant modifications to handle different languages. Dragon's software is available for American English, British English, French, German, Italian, Spanish and Swedish. Typing Chinese characters on computer keyboards is time-consuming because it can take six keystrokes to enter a single character. China is thus seen as a big potential market for speech recognition.
L&H recently unveiled a version of its software to handle continuous speech in Cantonese and convert it into traditional Chinese characters. The company also plans to launch two Mandarin language versions. IBM also has a Chinese version of its ViaVoice speech recognition software.
AT&T has started a research project to develop an English-Chinese speech translation system with a vocabulary of several thousands words. Intel, the world's biggest chip producer, is also keen on the Chinese market and last year hosted a conference in Beijing on speech
recognition.
Intel's interest in speech recognition stems from the fact that the algorithms that process speech are computer-intensive. If PC- based speech recognition does become popular, Intel reasons that users will want to upgrade to the latest PC hardware, such as Intel's Pentium III processor, announced last month. It has built into the Pentium III several new "extensions" to improve speech recognition. Dragon has developed speech recognition software that is optimised for the Pentium III and will be used to update existing products, such as Dragon NaturallySpeaking.
In spite of the interest in popularising voice recognition on PCs, it is unlikely to be the biggest market for the technology. The telephone is the only truly universal communications tool, so many companies are focusing on bringing voice recognition to a mass market via the phone.
"I believe the telephone is going to be the biggest application for voice technology," says Mr Patterson of SpeechWorks. It has been developing phone-based applications for online brokers, banks, travel agents and telephone companies. E*Trade, the big US online broker, has added mutual fund trading to the growing range of investment services that customers can access by phone using SpeechWorks software. E*Trade has offered a voice recognition service since 1997 for customers who cannot or do not want to use the internet to access their account or conduct a transaction. "Contrary to what they believe in Silicon Valley, not every-one is using the internet," says Mr Patterson.
The drive to reduce the often high staff costs of a call centre is causing more service organisations to consider voice recognition to automate routine transactions. The Gartner Group believes many healthcare,telecommunications and financial services companies will introduce speech recognition in the next two years. Seguros Comercial America, one of Mexico's largest insurers, plans to use software from Nuance, another US pioneer, to allow customers to access information and complete transactions by simply speaking their requests. Nuance has a version of its software for Latin- American Spanish speakers. Another Nuance user is Odeon Cinemas in the UK, which plans to let film-goers obtain information by speaking into the phone.
The ultimate goal of the industry is to marry speech recognition to the internet, thus opening e-commerce to the majority of the population that does not have an internet-equipped PC. More than a dozen companies, including Motorola, SAP of Germany, and Visa, teamed up with Nuance last year to launch an initiative, called V-Commerce, that will let consumers conduct net-based transactions by giving voice commands over the phone. The underlying key technology is Motorola's "voice mark-up language", called VoxML. By using special "voice browsing" software - which could be embedded in mobile phones, for example - words spoken over the telephone are converted into VoxML commands that can be sent over the internet and access information from any web server that understands VoxML. The reply would come back as VoxML commands and be converted into synthesised speech. And there lies the next challenge. While speech recognition has made tremendous progress, computer-generated speech remains quite crude. The reassuringly human-sounding words of Hal are still firmly in the realms of science fiction.
Copyright © The Financial Times Limited |
