MICHAEL鈥檚 office says he鈥檚 working at home today. You don鈥檛 know him very
well but you need to contact him urgently, so you dial his number. An
electronic-sounding voice bids you good morning and asks what you want. 鈥淓r, I
was hoping to speak to Michael,鈥� you reply. 鈥淲ho is calling?鈥� the voice asks. As
soon as you give your name, the voice is back: 鈥淚鈥檓 sorry, Michael has asked not
to be disturbed at present. Please leave a message after the tone.鈥� Before you
can say anything else, there is a whistling noise on the line.
Welcome to the world of the 鈥減ersonal agent鈥�, a technology that is just
around the corner, according to Bell Northern Research in Montreal. The agent
will sit at the exchange and field your calls, putting through only the people
you specify, and redirecting others to voice mail. The technology that will make
such agents possible is voice recognition.
This chilly brushoff may be yet to come, but computers that can pick out key
words and phrases and act on them are already popping up in telephone exchanges,
at the end of help lines, in hospitals and commercial companies. Directory
services in some American states are already operated by computer. With another
system, you say the name of a company over the phone and it tells you the latest
share price. And voice recognition software is making inroads as an alternative
to the keyboard as a way to communicate with PCs. Computers you can talk to were
a dream of the 1960s, but despite all the hype they have never lived up to
expectations. So why are they appearing now, and more importantly, becoming an
accepted part of everyday life?
Advertisement
According to Jackie Fenn, a research director with the advanced technologies
and applications service at the Gartner Group, one of America鈥檚 top business
research organisations, speech recognisers have now overcome critical cost and
credibility barriers. They have reached a point where organisations are prepared
to stake their businesses and reputations on them. Five years ago this was not
the case, she says. The first wave of systems sold in the early 1990s were often
unreliable, required users to leave long pauses between words, and were prone to
make wild guesses at words.
Today, although we are still nowhere near holding a natural conversation with
a computer, the systems have improved dramatically. For simple
tasks鈥攅specially over the phone鈥攖hey can give the same service as
people but for less money. The American phone company AT&T, for example, use
voice recognition to supervise a billion reverse-charge calls a year.
These advances are due in part to recent dramatic improvements in the power
of microprocessors. 鈥淎 couple of years ago people were using 386s. Today they
have 75 megahertz Pentium chips. That鈥檚 a massive change in processing power,鈥�
says Doug Sharp, a speech recognition researcher at AT&T Laboratories in
Murray Hill, New Jersey. At the same time, researchers have made small
refinements to the statistics and software used in speech processing and
recognition that add up to big gains. Some systems can even recognise limited
amounts of continuous speech.
These latter advances stem partly from the fact that almost all speech
recognition researchers are now taking the same basic approach to recognising
words. Five years ago, many university groups were trying to build speech
recognisers from neural networks鈥攖oday, most researchers have difficulty
thinking of anyone who is using just this approach. Instead, most of the
research effort is going into a single design based on a statistical approach to
identifying words. This needs less powerful computers and has proved to be a
much simpler approach than using neural networks or artificial intelligence
techniques.
Most of today鈥檚 statistically-based recognisers compare a person鈥檚 speech
against a vocabulary stored in memory. The first systems stored digital 鈥渕odels鈥�
of complete words, each of which contains details of how energy is distributed
across the frequency spectrum when the word is spoken. Modern systems use models
of phonemes, the smallest identifiable sounds in a language. English has 44, and
identifying these has proved faster than using whole words.
Once somebody鈥檚 spoken words have been digitised and the energy and frequency
data extracted, the results are compared to the models using a technique called
hidden Markov modelling (鈥淐omputers that Listen鈥�, New 杏吧原创, 4
December 1993, p 30). This calculates the probability that each model in memory
matches the spoken word. The system then 鈥渞ecognises鈥� the word with the highest
match probability.
Today, many speech recognition systems also include some form of language
model that can test words by using basic rules for grammar and context. As an
example, Melvyn Hunt of Dragon Systems UK, based in Cheltenham, quotes the
sentence 鈥淢r Wood would like to see the sea, but the wood would be in the way鈥�.
Dragon鈥檚 dictation software, which runs on a PC and allows people to talk rather
than type, 鈥渒nows鈥� that 鈥渨ould would鈥� is an unlikely combination and that 鈥淢r鈥�
is often followed by somebody鈥檚 name, so it would choose Wood beginning with a
capital letter as the most likely option.
One of the biggest challenges facing researchers is to make systems that can
recognise words spoken by people with different accents and pronunciations. The
early speech recognition systems were not good at this. As Fenn points out, one
of the problems with the first systems designed to work over the telephone was
that they could only understand American English. One system couldn鈥檛 understand
the word 鈥渙perator鈥� unless it was said in an American accent with American
pronunciation.
Speech recognition researchers have had some success in tackling this problem
by tweaking their systems. They now 鈥渢rain鈥� systems by asking thousands of
people with different accents to pronounce various words, and extracting the
energy distribution and frequency data for different phonemes. The 鈥渢olerance鈥�
of the comparison process can then be broadened to take account of all the
different ways of pronouncing the same phonemes. Although improved training has
made systems more robust, Sharp admits there are always some 鈥減athological
voices鈥� that are beyond the pale. 鈥淭hey鈥檙e a real problem,鈥� he says.
The other main obstacle facing speech recognition researchers is people鈥檚
desire to speak continuously without pausing between words. Most dictation
systems are 鈥渄iscrete utterance鈥� devices, says Hunt, people still have to leave
a slight pause between words. The problem facing researchers who want to make a
continuous speech recogniser is similar to one they have already solved to
improve recognition of individual words.
Sound work
Phonemes within a word change according to the preceding and following
phonemes. The sound made by 鈥渓鈥�, for example, changes between the words 鈥渟loop鈥�,
鈥渕ilk鈥� and 鈥渁nalyse鈥�. To overcome this problem, Dragon鈥檚 dictation system
actually focuses on three phonemes at a time鈥攐therwise known as phonemes
in context. In continuous speech, the problem is that the first and last
phonemes in a word can sound different depending on what comes before and after
them.
One way round this is to make digital models that bridge word boundaries. But
dictation systems, such as Dragon鈥檚, have active vocabularies of about 60 000
words, which would mean a lot of extra models. 鈥淭o do this you either pay a
heavy penalty in terms of memory or computational power, or you get clever,鈥�
says Sharp. 鈥淲e鈥檙e after the latter.鈥�
Some systems can already recognise continuously spoken phrases, such as
telephone services and systems like AT&T鈥檚 Watson, which allows you to
navigate round the applications on your PC by talking to it. They can store
models of entire phrases鈥攊ncluding word boundaries鈥攂y keeping their
vocabularies very small: typically between 16 and 60 words.
The inability of today鈥檚 dictation systems to handle continuous speech is a
major obstacle to their acceptance, according to Fenn. While computer users are
happy to learn how to type, they balk at having to change the way they speak in
order to use a speech recogniser. Makers of dictation systems, such as IBM a nd
Dragon, are racing to improve them so they can accept continuous speech.
At the same time, they want to reduce the amount of time that users must take
to 鈥渢rain鈥� the software to recognise their speech patterns and vocabularies.
When somebody buys a dictation system, it is programmed to accept a wide range
of pronunciations. But most systems can increase their accuracy and speed by
adapting to their user鈥檚 voice. They narrow down their statistical tolerance by
tuning in to properties such as the speed and pitch of a voice. The user also
adds words that the system does not know, so its vocabulary adapts.
Fenn has seen some reductions in 鈥渢raining鈥� times. IBM鈥檚 VoiceType dictation
system requires the user to spend just 90 minutes reading in words before it can
be used effectively, while the manual for Dragon dictation software claims no
training is needed. Hunt claims his company鈥檚 software adapts so quickly that
native speakers can use it straight out of the box without training it.
Fully-trained dictation systems have their advocates. Michael Jarmulowicz,
for example, a consultant pathologist at London鈥檚 Royal Free hospital uses IBM鈥檚
VoiceType software to dictate reports on tissue samples into his PC. His words
appear on screen, he prints them out and sends them back to the doctors who
ordered the analyses. He is quite happy with an error rate of about one mistake
in every four or five reports. Jarmulowicz believes that with speech recognition
he can turn a report round faster than when he relied on a secretary. Pathology,
of course, has its own language, with words such as 鈥渄yskariosis鈥� and 鈥渟quamous
cell carcinoma鈥�. To deal with these, IBM built a specialist vocabulary just for
pathologists.
Another group that Fenn says are benefiting from voice recognition are people
who cannot use their hands. Christopher Reeve, the actor who was paralysed
following a riding accident last year, finds his PC and speech recognition
system a great help. 鈥淚 am able to make phone calls, fax, and write letters,鈥� he
said recently. 鈥淚t really is one-stop-shopping. It keeps me in touch with the
飞辞谤濒诲.鈥�
But Fenn does not advocate that every professional person with a stack of
letters to write runs out and buys one of these systems. She believes that even
with all the improvements, typing and using a mouse is generally, faster, easier
and more effective for everyday purposes. But this is likely to change if the
processing speeds of microchips continue to rise, enabling more phonemes to be
checked in less time. Also researchers reckon that they will solve the problem
of recognising continuous speech before long, although they鈥檙e keeping their
methods close to their chests.
At present, there are still big problems to be overcome. Telephone
recognition systems are easily confused, for example. One researcher gives the
example of an experimental electronic timetable. If you tell it you do not want
the times of trains between two cities, it will tell you them just the same.
Also, StockTalk, an experimental system developed by Bell Northern Research to
quote company stock prices, is not 100 per cent accurate. Give it a fictitious
company name and it has been known to make a best guess at the name which is not
always correct.
Researchers from Carnegie Mellon University in Pittsburgh and the University
of Karlsruhe in Germany, believe that some of this confusion may be overcome by
combining speech recognition with hand writing and gesture recognition. So a
person could speak to their computer diary confirming an appointment, and at the
same time, block out the time for the meeting on a computerised calendar. A
change of mind would be signalled with a shake of the head. But since analysts
such as Fenn believe that these systems are still science fiction.
Hands off
Bell Northern Research is planning for the more immediate future. It has
developed telephone operator systems and 鈥渉ands-off dialling鈥�, in which you
simply say the name of the person you want to reach into the telephone and a
computer at the exchange dials the number for you. The company also reckons that
we will soon be talking to, rather than reading, Yellow Pages. A voice
synthesiser will ask what service you want, a recognition system will take your
order and hand back to the synthesiser to list relevant suppliers in your
area.
The Gartner Group also predicts that organisations such as stockbrokers are
likely to begin to introduce speech recognition for their traders鈥攚hose
vocabularies on the trading floor are usually quite small. Then there are
systems for department stores, which will ask what you want to buy and put you
through to the relevant department. For simple queries, the savings from using
computers rather than people could be enormous.
Already, companies are installing electronic operators which remove the
frustration of trying to leave a message for somebody after hours. You simply
say the person鈥檚 name and get switched through to their voice mail. But that
raises another question: do customers really want to be answered by a computer?
Isn鈥檛 voice mail bad enough?
The calls may be answered faster, but if you鈥檙e unfortunate enough to have a
鈥減athological voice鈥�, hard luck. And even if you don鈥檛, a confused computer may
try and give you an answer that bears no relation to your question. And you
can鈥檛 say to a speech recogniser, 鈥淗ang on a minute, I can smell something
burning in the oven鈥�, and expect it to still be there when you return.
Hunt, who is designing a speech recogniser for telephone systems, believes a
life line is essential. 鈥淵ou should be able to bale out of the system,鈥� he says.
鈥淚f you want an operator you should be able to say `operator鈥� and get one.鈥�