ad.nnet and ad.desc provide for a frame-based alphabet/digit
recognizer.  The pronunciations which should be used (this is how it
was trained and these categories are in the net) are listed below.

nnet.phlet is the neural network for the second pass re-evaluation
(libabc).  This net was trained on segmentations from ad.nnet and
ad.desc and should be used with it.  It has about 2/3 the error rate
for alphabet and the same error rate for digit.  However, the libabc
package can efficiently search a large data base which the frame-based
system could not do (efficiently).

Some grammars are provided here for convenience.  Others could be used
as needed.  For mixed alpha/digit strings, the system seems especially
susceptible to insertion errors and the grammar ad.hand_grammar
provides for insertion penalties to prevent this (used just like the
other kind of grammar but with the function call buildGrammarByHand).
 
a  [?] ei
b  [bc] b i:
c  s i:
d  [dc] d i:
e  [?] i:
f  [?] E f
g  [dZc] dZ i:
h  [?] ei tSc tS
i  [?] aI
j  [dZc] dZ ei
k  [kc] kh ei
l  [?] E l
m  [?] E m [&]
n  [?] E n [&]
o  [?] oU
p  [pc] ph i:
q  [kc] kh j u
r  [?] A 9r
s  [?] E s
t  [tc] th i:
u  [?] j u
v  v i:
w  [dc] d ^ bc [b] [l] j u
x  [?] E kc ks
y  w aI
z  z i:
zero  z I 9r oU
one  w ^ n [&]
two  [tc] th u
three  T 9r i:
four  f >r
five  f aI v
six  s I kc ks
seven  s E v & n [&]
eight  [?] ei [tc] [th]
nine  n aI n [&]
sil  .pau
any  .garbage
 
