PocketSphinx loading time

[Anonymous]

Hello, my english is not very good, sorry for that. I will try to explain
myself. First a small briefing:

Platform: Android + PocketSphinx.
Language Model Parameters: voxforge_es_sphinx.cd_cont_1500 - v.0.1.1.
Dictionary: 244 Words.
Grammar: JSGF.

When I started working with Android + PocketSphinix the Decoder takes about
one/three second to load all stuff using "hub4wsj_sc_8k" and that was great.
Now I am using the VoxForge Language Model Parameters for spanish and I am
getting about five/seven seconds at time of load the Decoder. That is too much
time.

I know that my dictionary and JSGF file loads very quickly. So the problem is
inside the Language Model Paremeters.

Question is: why those differences between loading time?

Thanks in advance.

Regards,

Lilo.

[eliasmajic]

Not really an answer but you only need to encounter that 5s/7s once. One thing
that I did is just hide that 5s in the loading screen they initially see when
running the app.

The spanish voxforge audio is also available to train your own model.

[Anonymous]

@eliasmajic thank you very much for your reply. By the nature of the
application, I can not wait the initial seven seconds. I know that I can add
all heavy load in a service executed at system boot and avoid the loading in
each launch of the application (this will be my last alternative).

About to train a new model I have doubts. Actually I am creating the
Dictionary and the JSGF in a dinamic way. So I don't know how to train a model
without a static content. Due to my bad english... my knowledges about that
are very poor. For that reason I like to know why one model is loaded faster
than other.

One more time, sorry for my bad english. I am not an smart monkey! I am an
average human :)

[Anonymous]

At this time, just with testing purposes I am trying to train a new Acoustic
Model using VoxForge stuff for spanish. But I'm getting the following Warnings
and Errors:

...
WARNING: This word: Té was in the transcript file, but is not in the dictionary (<s> TODO ESTABA LO MISMO QUE UNA HORA ANTES CUANDO EL Té HUMEABA EN LA TAZA DE OJEDA AHORA VACíA Y BLANQUEABAN SOBRE LA MESA LOS PLIEGOS </s> ). Do cases match?
WARNING: This word: VACíA was in the transcript file, but is not in the dictionary (<s> TODO ESTABA LO MISMO QUE UNA HORA ANTES CUANDO EL Té HUMEABA EN LA TAZA DE OJEDA AHORA VACíA Y BLANQUEABAN SOBRE LA MESA LOS PLIEGOS </s> ). Do cases match?
WARNING: This word: SALóN was in the transcript file, but is not in the dictionary (<s> DEL FONDO DEL SEGUNDO SALóN LLEGABAN CONFUNDIDOS CON RISAS DE MUJERES Y CHOQUE DE BANDEJAS </s> ). Do cases match?
WARNING: This word: FRíO was in the transcript file, but is not in the dictionary (<s> Y EL FRíO RESPLANDOR DE LAS AMPOLLAS ELéCTRICAS DESCENDíAN GORJEOS DE PáJAROS </s> ). Do cases match?
...





...
MODULE: 40 Build Trees
    Phase 1: Cleaning up old log files...
    Phase 2: Make Questions
    Phase 3: Tree building
        Processing each phone with each state
        A 0 
        A 1 
        A 2 
        B 0 
        B 1 
        B 2 
        CH 0 
        CH 1 
        CH 2 
        D 0 
        D 1 
        D 2 
        E 0 
        E 1 
        E 2 
        F 0 
        F 1 
        F 2 
        G 0 
        G 1 
        G 2 
        GN 0 
This step had 1 ERROR messages and 1 WARNING messages.  Please check the log file for details.
        GN 1 
This step had 1 ERROR messages and 1 WARNING messages.  Please check the log file for details.
        GN 2 
This step had 1 ERROR messages and 1 WARNING messages.  Please check the log file for details.
        I 0 
        I 1 
        I 2 
        J 0 
        J 1 
        J 2 
        K 0 
        K 1 
        K 2 
        L 0 
        L 1 
        L 2 
        LL 0 
        LL 1 
        LL 2 
        M 0 
        M 1 
        M 2 
        N 0 
        N 1 
        N 2 
        O 0 
        O 1 
        O 2 
        P 0 
        P 1 
        P 2 
        R 0 
        R 1 
        R 2 
        RR 0 
        RR 1 
        RR 2 
        S 0 
        S 1 
        S 2 
        T 0 
        T 1 
        T 2 
        U 0 
        U 1 
        U 2 
        X 0 
        X 1 
        X 2 
        Y 0 
        Y 1 
        Y 2 
        Z 0 
^C        Skipping SIL
MODULE: 45 Prune Trees
    Phase 1: Tree Pruning
This step had 1 ERROR messages and 0 WARNING messages.  Please check the log file for details.
    Phase 2: State Tying
This step had 1 ERROR messages and 0 WARNING messages.  Please check the log file for details.
MODULE: 50 Training Context dependent models
    Phase 1: Cleaning up directories:
    accumulator...logs...qmanager...
    Phase 2: Copy CI to CD initialize
This step had 1 ERROR messages and 0 WARNING messages.  Please check the log file for details.
    Phase 3: Forward-Backward
        Baum welch starting for 1 Gaussian(s), iteration: 1 (1 of 1)
        0% 
This step had 2 ERROR messages and 0 WARNING messages.  Please check the log file for details.
Only 0 parts of 1 of Baum Welch were successfully completed
Parts 1 failed to run!
Training failed in iteration 1
MODULE: 60 Lattice Generation
Skipped:  $ST::CFG_MMIE set to 'no' in sphinx_train.cfg
MODULE: 61 Lattice Pruning
Skipped:  $ST::CFG_MMIE set to 'no' in sphinx_train.cfg
MODULE: 62 Lattice Format Conversion
Skipped:  $ST::CFG_MMIE set to 'no' in sphinx_train.cfg
MODULE: 65 MMIE Training
Skipped:  $ST::CFG_MMIE set to 'no' in sphinx_train.cfg
MODULE: 90 deleted interpolation
Skipped for continuous models
...

Initially, Dictionary and Transcription files was in UTF-8 but I was getting
"broken" characters in the log output.. so I set both files to ISO-8859-15.
Then all characters like áéíóú was showed fine, but I have the same errors.

[Nickolay V. Shmyrev]

WARNING: This word: VACíA was in the transcript file, but is not in the
dictionary ( TODO ESTABA LO MISMO QUE UNA HORA ANTES CUANDO EL Té HUMEABA
EN LA TAZA DE OJEDA AHORA VACíA Y BLANQUEABAN SOBRE LA MESA LOS PLIEGOS
). Do cases match?

Dictionary might require resorting after you converted it to another encoding.
It's recommended to use UTF-8 , not 8859-15.

The warning just tells you that the word is missing. All you need to do is to
make sure word is there.

[Anonymous]

Thank you for your reply. Initially I was using the original file (encoded
with UTF-8) but I got the following warnings:

...
WARNING: This word: SEñALA was in the transcript file, but is not in the dictionary (<s> SEñALA CON EL DEDO LO QUE MáS TE GUSTE </s> ). Do cases match?
...

It should be:

...
SEñALA instead of SEñALA
MáS instead of MáS
...

Dictionary extraction:

...
SEñALA     S E GN A L A  
...

Transcription extraction:

...
SEñALA CON EL DEDO LO QUE MáS TE GUSTE (ciego-10122009/wav/78)
...

[Nickolay V. Shmyrev]

This seems to be a mistake in the original voxforge data. I think it's easy to
fix it in the archive itself and then train.

[Anonymous]

I think that both (Dictionary and Transcription) files are not mistaken in the
original VoxForge, I can read them without broken characters. How can I double
check it?

Please, sorry for by bad english.. maybe I misunderstood something. Thank you!

[Nickolay V. Shmyrev]

Maybe the training script that creates prompts using awk breaks the text. You
might want to set locale on your machine to en_US.UTF-8 or to C. It could
affect awk behaviour.

[Anonymous]

Same result as before. But don't worry I don't like to bother you.

Just one thing: the idea to create a new acoustic model is due to I am getting
a "Decoder" load time of seven seconds using PocketSphinx + Android (please
check my post number 1 and 3). But I think that it will be the same, due to it
will be only a new compilation using the same base. So... it is there any way
to reuse the current VoxForge in Spanish doing it like "hub4wsj_sc_8k" that
takes only 3 seconds to load? Maybe I am asking silly questions...

~$ locale -a
C
C.UTF-8
de_CH.utf8
en_AG
en_AG.utf8
en_AU.utf8
en_BW.utf8
en_CA.utf8
en_DK.utf8
en_GB.utf8
en_HK.utf8
en_IE.utf8
en_IN
en_IN.utf8
en_NG
en_NG.utf8
en_NZ.utf8
en_PH.utf8
en_SG.utf8
en_US.utf8
en_ZA.utf8
en_ZM
en_ZM.utf8
en_ZW.utf8
es_AR.utf8
es_BO.utf8
es_CL.utf8
es_CO.utf8
es_CR.utf8
es_DO.utf8
es_EC.utf8
es_ES.utf8
es_GT.utf8
es_HN.utf8
es_MX.utf8
es_NI.utf8
es_PA.utf8
es_PE.utf8
es_PR.utf8
es_PY.utf8
es_SV.utf8
es_US.utf8
es_UY.utf8
es_VE.utf8
POSIX
zh_CN.utf8
zh_SG.utf8
~$ LC_ALL="en_US.utf8"
~$ locale
LANG=en_US.UTF-8
LANGUAGE=en_US:en
LC_CTYPE="en_US.utf8"
LC_NUMERIC="en_US.utf8"
LC_TIME="en_US.utf8"
LC_COLLATE="en_US.utf8"
LC_MONETARY="en_US.utf8"
LC_MESSAGES="en_US.utf8"
LC_PAPER="en_US.utf8"
LC_NAME="en_US.utf8"
LC_ADDRESS="en_US.utf8"
LC_TELEPHONE="en_US.utf8"
LC_MEASUREMENT="en_US.utf8"
LC_IDENTIFICATION="en_US.utf8"
LC_ALL=en_US.utf8
~$ cat /etc/default/locale
LANG="en_US.UTF-8"
~$ reboot

Thank you Mr. Nickolay.

[Nickolay V. Shmyrev]

Just one thing: the idea to create a new acoustic model is due to I am
getting a "Decoder" load time of seven seconds using PocketSphinx + Android
(please check my post number 1 and 3). But I think that it will be the same,
due to it will be only a new compilation using the same base. So... it is
there any way to reuse the current VoxForge in Spanish doing it like
"hub4wsj_sc_8k" that takes only 3 seconds to load? Maybe I am asking silly
questions..

.

As said above, you need to train semi-continuous model from the Voxforge data.
Also you can convert mdef file to binary format with
pocketsphinx_mdef_convert. Then it will load faster.

[Anonymous]

Thank you! I am already using a mdef file in binary format. There is about one
second difference with the mdef text file. So, I will try to create a new
model using a different computer.

Really, thank you very much for your time.

I was trying to edit my first post to put everything in order, but I can not.
I will maintain you updated with my progress.. if you need something from me,
just poke me :)