Pocketsphinx lexical tree.

[creative64]

Hi,

I've a simple grammar with two words GENERAL and SILENT.

My Grammar file looks like:

Public = GENERAL | SILENT;

My dictionary contains pronunciations for these words as follows:

GENERAL JH EH N ER AH L
GENERAL(2) JH EH N R AH L
GENERAL(3) JH AH N R AH L
SILENT S AY L AH N T
SILENT(2) S AY L EH N T

When I look at the lexical tree created for this, I find that lexical tree
for state 0 is created something like this (filler omitted for clarity).

JH->EH->N->ER->AH->L
|
JH->EH->N->R->AH->L
|
JH->AH->N->R->AH->L
|
S->AY->L->AH->N->T
|
L->EH->N->T

I would have expected it to share initial nodes between multiple
pronunciations
of GENERAL as it does for SILENT.

When I remove the third pronunciation of GENERAL by changing the dictionary to

GENERAL JH EH N ER AH L
GENERAL(2) JH EH N R AH L
SILENT S AY L AH N T
SILENT(2) S AY L EH N T

Things become as expected and the lexical tree comes out as follows:

JH->EH->N->ER->AH->L
| |
| N->R->AH->L
|
S->AY->L->AH->N->T
|
L->EH->N->T

I'm using hub4wsj_sc_8k acoustic model and pocketsphinx version prior to 7.

Is this deliberate or an optimization issue ?

Thanks and regards,

[Nickolay V. Shmyrev]

Lextree is a tree of triphones, not a tree of phones. So your graph is not
really relevant to pocketsphinx.

[creative64]

Hi Nickolay,

Triphone details are embedded in this diagram. A phone shown in the diagram
can be thought as triphone which is
composed of this phone surrounded by its left and right context as per the
graphs. For root and leaf nodes, the left and
right contexts are SIL (Basically this graph only shows ci_ext field of
fsg_pnode_s structure for lextree).

These graphs are generated from the actual lextree dump taken from
pocketsphinx (triphone information was manually
checked by printing senone id sequences for each node and was found to be
matching with the graph).

Thanks and regards,

[creative64]

just to add "->" shows successor node and " | "shows sibling.

[Nickolay V. Shmyrev]

Triphone details are embedded in this diagram.

If you properly draw triphones your question will disappear I think.