home
***
CD-ROM
|
disk
|
FTP
|
other
***
search
/
ftp.whtech.com
/
ftp.whtech.com.tar
/
ftp.whtech.com
/
articles
/
The Economist
/
Chips that speak and hear.txt
next >
Wrap
Text File
|
2007-12-19
|
3KB
|
48 lines
The Economist
August 22, 1981
Chips that speak and hear
SECTION: Business, finance and science; SCIENCE BRIEF; Pg. 80 (U.S. Edition Pg. 72)
"Here is the news"
So far, most speech-synthesis chips have gone into electronic toys and learning
aids. Incorporating them into arcade games like Space Invaders is this year's fad.
Some have also been purchased by companies experimenting with speech applications
for cars, cookers and automated bank tellers. One application stands out as the
biggest challenge yet for speech-synthesis experts: the text-to-speech translator--
in effect, a reading machine. One such translator which will be on the market in
September can be attached to Texas Instruments' home computer, the TI 99/4. The
translator, which will cost around $200, reads aloud any news or information
displayed on the computer's television screen.
Two steps are usually involved in converting text to speech. First, special rules
are followed to translate the letters of the text into binary numbers representing
component sounds. The components can be the sounds of complete words or phrases,
bits of words or variations of individual letter sounds. Second, the components
stored as binary numbers are strung together in the right groupings and sequences
depending on syntax. In TI's approach, the component sounds used are finely
differentiated. They come from a library of 128 sound elements called allophones.
The ''p'' in push and the ''p'' in Spain are different allophones of the phoneme
''p''. Even so, the same words can sound wrong when uttered in a different context.
An even more ambitious reading machine has been developed by Professor Jonathan
Allen at the Massachusetts Institute of Technology. Called Mitalk, it can read not
only unrestricted English text or any word currently in the English language, but
also any future construction or even nonsense--including the poem Jabberwocky.
Mitalk's sound components are 7,000 root words, prefixes and suffixes called morphs
(eg, scarcity = scarce + ity). It produces intelligible and rather machine-like
speech, but is prohibitively expensive for most applications because it needs
600,000 bytes of storage space for its morphs (one byte, equal to eight bits of
data, is required to define each letter). By contrast, the even less natural-
sounding TI text translator needs only 3,000 bytes of memory for its allophones.
Professor Allen says Mitalk is modelled on a flat midwestern American accent, and
that it would take years to make it speak any other. What must Matsushita, a Mitalk
licensee, make of this limitation? A small Californian company called Telesensory
Systems is trying to make a scaled-down version of Mitalk as an aid for the blind.