- This topic has 15 replies, 5 voices, and was last updated 11 years, 4 months ago by .
-
Posts
-
I’m trying to implement an app that reads letters and numbers. I’ve had some inaccuracies for certain letter combinations—“AB” is consistently heard as ”8Y,” for instance, and am wondering if there are any configuration options that might help. I’ve had some success replacing letters with equivalents: ‘b’ with “bee,” etc. I have had much more success using the NATO alphabet (alpha, beta, etc.), but we can’t expect our users to be able to use it. So… what’s the best way to approach single letters? Thanks in advance.
Does the app read letters and numbers or recognize them in the user’s speech?
Would it be possible for you to show me your language model?
Can I also see the dictionary? I’m surprised to hear that it is recognizing EIGHT Y for A B; the EIGHT isn’t surprising but the Y is. Is EIGHT Y an accurate transcription of what Pocketsphinx heard? What is the hypothesis (verbatim)?
OK, I think the first easy step is to get rid of the pronunciations that are in the dictionary that you definitely don’t want to recognize. I realize this isn’t at all self-evident so I’ll explain briefly. If you look at this from the dictionary:
A AH
A(2) EYThat means that the language model tool gave you back two possible pronunciations for the word A. The first one is the particular NA pronunciation of the article “a” as in “a dog barked” that rhymes with “huh”. Since you don’t ever want to recognize that pronunciation of “a” because the alphabet character is never pronounced that way, you should erase that pronunciation from your dictionary.
The (2) in parentheses just means that it is the second pronunciation of the word, so the way you would want to replace
A AH
A(2) EYis with the line
A EY
deleting the first pronunciation, and removing the (2) from the second pronunciation since it is now the only pronunciation you are going to accept.
The next thing that you can do is to make the sentence “A B” part of your corpus. The corpus can have individual words, but it can also contain combinations of words. Combinations of words that you have made part of your corpus will have an automatically higher probability of being detected.
So, the corpus would say something like this:
A
B
A B
C
D
E
F
G
H
I
J
K
L
M
N
O
P
Q
R
S
T
U
V
W
X
Y
Z
ONE
TWO
THREE
FOUR
FIVE
SIX
SEVEN
EIGHT
NINE
ZEROYou can do this for all of the possible combinations if you want to, or just the ones where you want to raise their probability of being detected. When you look at the language model that is output, you will see that there is a 2-gram entry for A B and that it has a raised probability.
Just as an update on this, I’ve been gradually learning that individual letters/syllables are a challenging case and expectations for accuracy should probably be lower than for whole word or phrase recognition.
Letter recognition can only be done in conjunction with a spelling application. In other words, if you have a list of street names, spelling
W O O D W A R D
will work, if you use an n-best list and search through the dictionary for the results. As long as your task can be constrained by a dictionary, even a huge dictionary, you’re OK. You’ll have to patch OpenEars for N-best output, though, and build an FSG or LM. The LM will work better if you build it from your dictionary.
If your letter sequences truly are random, you’re dealing with something that’s beyond the state of the art. It’s beyond the state of the art for human listeners, too. Give it a try, read some random letter sequences to people and see how many they get wrong.
Is this something you’re still working on?
I am working on a similar project and having the same issues. As a matter of fact, recognition results are very poor. The biggest problems I am having with letters E, D, P, A. Below is the dictionary file that I compiled. In many cases it recognizes E as P, D, C. Letter A recognizes as 8 or H.
What techniques would you recommend for improving accuracy? Doesn’t more combinations for the same letter help?
0 Z IY R OW
1 W AH N
2 T UW
3 TH R IY
4 F OW R
5 F AY V
6 S IH K S
7 S EH V AH N
8 EY T
9 N AY N
A AH
A(2) EY
B B IY
C S IY
D D IY
E IY
E(2) EH
E(3) IH
F EH F
G JH IY
H EY CH
H(2) EY JH
I AY
J JH EY
K K EY
L EH L
M EH M
M(2) AE M
N EH N
N(2) AE N
O OW
P P IY
Q K Y UW
R AA R
S EH S
S(2) AE S
T T IY
U Y UW
V V IY
W D AH B AH L Y UW
X EH K S
X(2) AE K S
Y W AY
Z Z IYThis is not a good application of the library, unfortunately.
“As long as your task can be constrained by a dictionary”
Joseph, is there a way we can do this? e.g. only recognise “sentences” e.g. “w o o d w a r d” and not individual words e.g. “w”
I’m attempting to create such a spelling app
- You must be logged in to reply to this topic.