|
|
Preface
|
|
|
|
|
|
vocabulary was restricted, and the output speech quality required extensive learn-
|
|
|
ing. At that time, the computer implementation used for research consisted of a
|
|
|
Digital Equipment Corporation PDP-1 used for character recognition and morph
|
|
|
analysis, which was coupled to MIT Lincoln Laboratory’s TX-O computer (the
|
|
|
only one of its kind) for the synthesis algorithms. T. P. Barnwell III and
|
|
|
E. R. Jensen were responsible for building much of this computational environ-
|
|
|
ment. This required great effort and coordination, since all coding was performed
|
|
|
in assembly language.
|
|
|
|
|
|
Following the late 1960s, the character recognition and speech synthesis ef-
|
|
|
forts continued independently of one another with the work of B. Blesser,
|
|
|
M. Eden, and D. E. Troxel focused on the character recognition efforts. J. Allen
|
|
|
joined the faculty in September, 1968. Goals for a fundamental and comprehen-
|
|
|
sive research program aimed at the computation of high-quality speech using un-
|
|
|
restricted English text as input were formulated. In addition, strong coupling con-
|
|
|
tinued with the Speech Communication and Linguistics groups within the
|
|
|
Research Laboratory of Electronics, led by K. N. Stevens and M. Halle, respec-
|
|
|
tively.
|
|
|
|
|
|
With the desire to convert unrestricted English text to speech, a new scheme
|
|
|
was developed for the pronunciation of all possible English words. This required
|
|
|
elaborate extensions to the morph decomposition process, as well as the construc-
|
|
|
tion of a comprehensive morph lexicon to serve the entire language. Furthermore,
|
|
|
spelled speech was rejected as inadequate, and plans for the development of a
|
|
|
comprehensive set of letter-to-sound rules that would complement the morph
|
|
|
analysis process were established. In order to build a new morph lexicon, a copy
|
|
|
of the Brown corpus was obtained and sorted (shortest word first). Initial phonetic
|
|
|
segment labels were obtained from a computer-readable copy of the Merriam
|
|
|
Pocket Dictionary. Beginning with a nascent lexicon containing all bound morphs
|
|
|
and function words, each word from the Brown corpus was successively analyzed.
|
|
|
|
|
|
This led to the interactive addition of new morphs and a great deal of experience
|
|
|
with morph analysis procedures. This process was accomplished by J. Allen and
|
|
|
D. A. Finkel, with algorithmic and programming support from E. R. Jensen and
|
|
|
F. X. Carroll. The process spanned many months, and led to the extension of
|
|
|
morph analysis routines to include multiple decompositions and attendant selection
|
|
|
rules. The computational support for this work was a Digital Equipment Corpora-
|
|
|
tion PDP-9 computer with 24K words of memory and DEC tapes for peripheral
|
|
|
storage. Readers familiar with this equipment will have some appreciation of the
|
|
|
sheer magnitude of the effort required to build the morph lexicon and acquire the
|