"computing transitive closure&quo...

[Anonymous]

When doing a test speech recognition task with a FSG grammar (actually JSGF,
but converted to FSG, and JSGF is just as slow), the FSG grammar parsing takes
a really long time, and a lot of CPU power "Computing transitive closure for
null transitions". The full log is

INFO: fsg_model.c(264): 15597 null transitions added
INFO: fsg_model.c(411): Adding silence transitions for <sil> to FSG
INFO: fsg_model.c(431): Added 2573 silence word transitions
INFO: fsg_model.c(411): Adding silence transitions for ++NOISE++ to FSG
INFO: fsg_model.c(431): Added 2573 silence word transitions
INFO: fsg_model.c(411): Adding silence transitions for ++BREATH++ to FSG
INFO: fsg_model.c(431): Added 2573 silence word transitions
INFO: fsg_model.c(411): Adding silence transitions for ++SMACK++ to FSG
INFO: fsg_model.c(431): Added 2573 silence word transitions
INFO: fsg_model.c(411): Adding silence transitions for ++COUGH++ to FSG
INFO: fsg_model.c(431): Added 2573 silence word transitions
INFO: fsg_model.c(411): Adding silence transitions for ++LAUGH++ to FSG
INFO: fsg_model.c(431): Added 2573 silence word transitions
INFO: fsg_model.c(411): Adding silence transitions for ++TONE++ to FSG
INFO: fsg_model.c(431): Added 2573 silence word transitions
INFO: fsg_model.c(411): Adding silence transitions for ++UH++ to FSG
INFO: fsg_model.c(431): Added 2573 silence word transitions
INFO: fsg_model.c(411): Adding silence transitions for ++UM++ to FSG
INFO: fsg_model.c(431): Added 2573 silence word transitions
INFO: fsg_search.c(364): Added 211 alternate word transitions
INFO: fsg_lextree.c(108): Allocated 262446 bytes (256 KiB) for left and right context phones
INFO: fsg_lextree.c(251): 34111 HMM nodes in lextree (28786 leaves)
INFO: fsg_lextree.c(253): Allocated 4366208 bytes (4263 KiB) for all lextree nodes
INFO: fsg_lextree.c(256): Allocated 3684608 bytes (3598 KiB) for lextree leafnodes

Since it does this every time, is there a way I can convert it to a format
where this information is pre-computed? I noticed that DMP language models
start up very quickly, but I can't seem to find any way to convert from a FSG
to a DMP. With my FSG, it takes 30 seconds to do this step, and then like half
a second to do the actual recognition.

(My FSG grammar is not a "toy" grammar, but it is pretty simply. It's
basically just combinations of a few verbs with a few nouns, and a few filler
words mixed in. I know that it is reasonable to expect it to start up in a
small amount of time, becuase it is the direct translation of an HTK grammar
which started up in <half a second.

[Nickolay V. Shmyrev]

Please share the grammar

[Anonymous]

JSGF v1.0;

grammar edu.jas.adventure;

<weapon> = axe | wand | rod; 
<article> =(the | an | a);
<lamp_item> = lamp |lantern | light | headlight| (head lamp);
<ynanswer> = yes | yeah | okay| no | nope;
<item> = <weapon>| <lamp_item> | rock  | surface | battery | carpet | barrel | gold | ([gold] nugget) | diamond | 
([silver] bars) |silver | jewel | coins | chest | box | treasure | eggs | egg | nest | trident | ( [ming] vase)  | (shards [of pottery]) | pottery | emerald | platinum | 
pyramid | (platinum pyramid) | pearl | ([persian] rug) | spice | chain | 
([vending] machine) | knife | knives | food | rations | bottle | jar | water | (h two oh)| oil | debris | shell  | keys | key | grate | cage |  pillow | velvet | tablet | magazine | issue | 
mirror |  mist | lock | nothing | oyster | clam;

<organism> = troll | bear | dragon | bird | pirate | dwarf | dwarves | snake | clam | oyster | plant | 
giant | shadow | figure  ;

<voice_verb> = (say | chant | sing | utter | mumble) ;
<location> =  depression | road | hill | forest | valley | gully | stream | passage | tunnel | 
chasm | fissure | pit | outdoor | outdoors | crack | cave | entrance | slit | slab | 
 tree | trees | (bed quilt) | office | cavern | floor |
  house | dome | hall | view | staircase | stairs | steps | (oriental room) | reservoir | volcano | geyser|hole | wall 
 | message |(low room) | building |drawings|
     door | grate | fork | bed |moss |carpet| ([awkward|secret] canyon) | ([broken] rock) |
     (main (corridor | office)) |([debris | dark] room) |(low (window| passage | tunnel));
<move_action> = enter | return | jump | climb | retreat | retread | continue | walk | run | exit | leave | crawl | ascend | descend | cross|swim ;
<ambiguous_move_action> = travel | proceed | go | explore | follow | turn ;

<unambiguous_direction> = upstream | downstream | forward | onward | back | outside | inward | nowhere | downward |left | right | upward | east |  west | north  | south;

<explosive_verb> = blast | detonate | ignite | (blow up);
<attack_verb> = attack | kill | slay | fight | hit | strike | smash | <explosive_verb>;
<magic_verb> = abra | abracadabra | (hocus pocus) | (x y z z y) | (open sesame);
<item_verb> = carry 
       | (pick up) 
       | take | keep | catch | steal | capture | get | tote | drop | release | free | discard | dump | throw | toss | find | rub | fill | shatter | break | read | peruse;
<menu_action> =  (where (am i))
         | (i am lost) 
         | help | inventory | suspend | pause | save | stop | quit | hours | score |  info | information| brief;

<examine_verb>= (describe  | look | examine | touch) ;

<item_to_open_close> = door | grate |house | oyster | clam |chain | jar | bottle;
<open_close_verb>= (unlock | open | lock | close) ;
<digging_verb> =dig | excavate ;

<animal_verb> = feed | calm | tame | placate | wake | disturb;
<animal_action> = <animal_verb> [<article>] <organism> ;

<explosive_action> = <explosive_verb> [<article>] <location> ;

<target> =<organism> | <item>;
<fight_action> = <attack_verb> [[<article>] <target>] [with [<article>] <item>]  ;

<attack_item_verb>= wave | shake | swing;
<attack_item_action> =  <attack_item_verb> [<article>] <item> ;

<food_verb> = pour | eat | devour | drink;
<food_action>  = <food_verb> [<article>] <target> ;

<off_on> = (on|off);

<lamp_verb>    =light 
          |extinguish 
          | (turn [<off_on>]) ; 
<lamp_action> = (<lamp_verb> [<article>] <lamp_item>) 
         | (turn [<article>] <lamp_item> <off_on>) ;

<digging_action> =  <digging_verb> [<location>] ;


<item_action>= <item_verb> [<article>] <item>;

<magic_action> =[<voice_verb>] <magic_verb>;

<examine_target>=(<location> | <item> 
               | <organism>) ;

<examine_action> =<examine_verb> [[at] [<article>] <examine_target>] ;


<fluid_verb> =(water | oil);    
<fluid_action> =  <fluid_verb> [<article>] <target> ;

<open_close_action> =<open_close_verb> [<article>] <item_to_open_close> [with [<article>] <item>] ;


<opt_of>           =([of]);
<modifier_direction>= to | into | on | (off <opt_of>);
<opt_modifier_direction> = ((inside | out) <opt_of> )| in | up | above | down | across | over | under | onto;

<directional_phrase> = <unambiguous_direction> 
            | (<modifier_direction> [<article>] <location>)
            | (<opt_modifier_direction> [[<article>] <location>]) ;
<move_phrase> = (<move_action> [<directional_phrase>]) 
         | (<ambiguous_move_action> <directional_phrase>) ;

<thing_to_say> = <item> | <organism> | <location> | hello | hi | <ynanswer>;
<voice_action> = <voice_verb> <thing_to_say> ;


public <command> = <ynanswer> | <voice_action> | <move_phrase> | <open_close_action> | <fluid_action> | <food_action> | <menu_action> | <examine_action> | <magic_action> | <item_action> |<lamp_action> | <digging_action> | <fight_action> | <attack_item_action> | <explosive_action> | <animal_action>;

[Nickolay V. Shmyrev]

Hello

Thanks for the information. I've just committed a fix to speedup closure
computation. It should help, please update the sphinxbase.

AFAIK HTK doesn't do closure that explains it starts faster but works slower.