|
| |
Speech
Recognition has been of interest to scientists for more than a century. The
evolution of the study of Speech Recognition began with extensive research into
the complex biological nature of the human voice. Applications to these findings
on audible speech started to take form with the help of the computer. Today, SR systems are quickly becoming reliable and accurate programs
used to
recognize and interpret the spoken word.
Important Events in the
Evolution of Speech Recognition
 |
1874:
Alexander Graham Bell proves that frequency harmonics from an electrical signal
can be divided. This eventually leads to the digitization of speech.
|
|
1952:
Bell Labs develops the first effective speech recognizer (97% accurate) using
the
simple frequency splitter technology similar to the one developed by Alexander
Graham Bell 78 years ago.
|
|
1969:
Computer systems such as the Vicens System and the Medress System are developed with
limited vocabularies in memory. Intermittent speaking, or discrete
speech,
is required for the voice to be recognized.
|
|
1971-1976:
U.S. Defense Advanced Research Projects Agency (DARPA) funds projects for Speech
Understanding Research (SUR).
|
The major findings from this initiative
include
improvements to reduce the problem of
variability in the voice:
|
Japanese
scientists Itakura, Skoe, and Chiba create dynammic
programming which becomes the standard for optimal non-linear time
alignments.
|
|
Markov
Modeling by
Jim Baker and IBM’s Fred
Jelinek - a mathematical model for located invariant
information in the speech signal.
|
|
Mid
1970’s:
|
|
Itakura
conducts research and develops product on the basis that noises may sound
similar, but actually are different. His product tested 97.3% accurate with a
vocabulary of about 200 words.
|
|
Bell Labs
develops a system with a 97.1% accuracy rate that can interpret voices from
different people (however the vocabulary bank is very small).
|
|
Late
1970’s: commercial Speech Recognition packages become available. Prices range
from $259 to $100,000.
|
|
1980’s:
The SR market splits in two directions: 1) Call center Speech Recognition systems and
2) Speech-to-text applications.
|
|
1990’s:
Processing power of the personal computer reaches the necessary level for SR
software to be used effectively by ordinary user (about 200Mhz).
|
|
1999: SR
programs can understand continuous speech from multiple users with up to 99%
accuracy. The time needed for system to learn a user’s voice is under 10
minutes. Vocabulary databases exceed 250,000 terms and 160,000 active words.
|
|
