multiple Pronunciations

[tarekabuamer]

When I use multi Pronunciations for dictionary words in pocket sphinx it
always gives NULL as a recognition result, while if I add them as different
words they are recognized correctly. Any clue?

[Nickolay V. Shmyrev]

Sorry, it's hard to understand your short description.

Please provide an example to reproduce your problem.

Please provide more information. What pocketsphinx version are you using.

[tarekabuamer]

suppose I have two different Pronunciations p1 and p2 for word w. By default I
add them to the dictionary like this:

w p1
w(1) p2

(two Pronunciations for the same word as in sphinx format) in this case when I
utter p1, w is returned but when when I utter p2 NULL is returned by the
recognizer.

if I put them in the dictionary like this:

w p1
w1 p2

(two different words). when I utter p1 w is returned and when I utter p2 w1 is
returned.

My question is: why the NULL in the first case.
I am using pocketsphinx version 7 (the latest)

[Nickolay V. Shmyrev]

And what grammar or language model do you use

[tarekabuamer]

simple jsgf grammer:

JSGF V1.0;

grammar words;
public

<main> = w; //in first case of dictionary
public <main> = w | w1; //in second case of dictionary

</main></main>

[Nickolay V. Shmyrev]

Sorry, I can not reproduce your problem. You can get my test here

http://dl.dropbox.com/u/26073448/test-
multpron.tar.gz

[Anonymous]

Hi,

To add to this, I am seeing a similar issue with FSGs, when querying the
hypothesis during recognition. What I see is hypothesis that should not be
allowed from the grammar, basically missing a word. It only happens on words
that have several pronunciation alternatives. I have been trying to set up a
reproduce-able test case using pocketsphinx_continuous, but failed to
reproduce it so far.

I do not see this issue in the final recognition (after 2nd pass), and do not
see it if all words have only one pronunciation.

I'll post again if I can set up an environment to reproduce the issue.

Sylvain

[Joseph S. Wisniewski]

It's the probably the (1).

Sphinx requires the alt prons to start at 2.
w p1
w(2) p2

Also, make sure your editor is not expanding tabs into spaces (thanks loads,
Apple!) because spaces are parts of words.

w(2)<space>p2 creates a word "w(2) p2" without a pron
w(2)<tab>p2 creates an alt pron for w, with the pron p2.</tab></space>

[Joseph S. Wisniewski]

Aargh. Brain freeze. Scratch that. Starting with (2) is just "tradition",
because the first pron doesn't get a number. It doesn't care if the alts start
with (1) or (2).

[Joseph S. Wisniewski]

Aside from checking for spaces, I'm running out of ideas. Two things, but
they're longshots.

Make sure you're not changing the -fsgusealtpron parameter (the default is
"yes" already)

Watch your recognizer output and see if you get something like this in the
output:

INFO: fsg_model.c(325): Adding alternate word transitions (word,word(1)) to
FSG
INFO: fsg_model.c(358): Added 1 alternate word transitions
INFO: fsg_search.c(364): Added 1 alternate word transitions