VoxForge
VoxForge
Hi
I
think now in theory it should work. I installed sphinx and the shinx
backend is compiled in. The adapted model does compile fine and no erros are reported.
I use the german model from guenter with adaption. Shadow dict from ralfherzog.
I dont think this is a language/dict problem so i post it in general.
However:
It does not seem to recognise anything!! I have trained the word
firefox over 130 times and it is also in the shadowdict and i guess it
is in the base model also.
So there should be some basis but is does not regognise anything. I dont belive that im repeatetly capeable of more than 130 ways of speaking firefox, so ther must be an issue somewhere.
Some other info who is might helpful:
Voice-Activity-Detection -Standart values
SNR: 150 with should be like 1,5 x my noise base level with seems sine in thery as my baste noise level is around 7 % of mic rec vol and i speak like max 75 % of mic rec vol and start by 16 or so. Before it was 2300 with would be 23x like nise level, that is ridiculus with my nose level here and low tech mic....
But this could be wrong i guess.
Also i did have a noise filter post processing who would remove my backgroundnoise but then i thought, when i speak live i doubt then these postprocessing filter will be used so there is a huge difference between those files and the training data without nose and now deactived it.
It would also be a help to know, how many recordings do you normaly need to get somehting with simon to work?
If you need more info, please don't hesitate to ask :-D
Thanks a lot guys :-D
--- (Edited on 3/25/2014 4:51 pm [GMT-0500] by Manu) ---
> I use the german model from guenter with adaption. Shadow dict from ralfherzog.
Those are not compatible I think. You need to use the dictionary from the model.
> I dont think this is a language/dict problem so i post it in general.
It is a language/dict problem
> It would also be a help to know, how many recordings do you normaly need to get somehting with simon to work?
--- (Edited on 3/26/2014 01:15 [GMT+0300] by nsh) ---
Ok, i downloaded the dict who is supplied with the model.
so:
acustic model + shadow dict do match now i guess.
But what is with the scenario(s) "Window management" it does bring the pronounication of certain words with it.
How do i know what dict was used to make them ?
Do i need to replace them with the pronounication of the shadow dict ? I would guess so. They do not match.
EDIT: I used the new pronounciation and now the word is red marked. I read, that this means the pronounciation does not match the one from the model.
SO basicly we are where we left of. Nothing is recognised and the dicts seems to be the right one. Right ???
EDIT":
Also i installed HTK as this is written in the manual would be used for adapted models. But i'm confused here, does not sphinx provide the same functionality ?
With the installation of htk, i purged everything. So no training data anymore.
I will now try again to use the german base model, this time with the voxforge.dict inside of the model and then try to train "firefox"
For now (54 trainings + shadow dict from model) this is not working....
Also it says in many tutorials that the vocaulary words would be red if no traing data is available. This is also not the case here.
Thanks
--- (Edited on 3/25/2014 5:51 pm [GMT-0500] by Manu) ---
--- (Edited on 3/25/2014 6:02 pm [GMT-0500] by Manu) ---
--- (Edited on 3/26/2014 5:43 am [GMT-0500] by Manu) ---
--- (Edited on 3/26/2014 6:00 am [GMT-0500] by Manu) ---
Hey guys :-D
I played around some more and found some things out:
I trained the words with simon, exported the model and tried it out with pocketsphix.
Result: pocketsphinx hast nearly 100 % recognition rate.
For me this means the follwing:
- Training with simon works and the changes are actually reflected in the accustic model.
- Only the actual recognition does not work, but all the backends are ok i guess and it should be possible to use it if i would get Simon to process the audio it records.
- So a misconfigured Simon is more likely the reason. (Maybe?)
- It would also mean that simon does use sphinx and not pocketphinx to recognise my spee right ?
-It also means, that if simon does not "want" my speech i could use any other programm to control my PC with.
But i dont know any other expect from blather but i don't think this would work with the german model.
Do you know any of those programs ? Or why/how i maybe have misconfigured simon ?
Thanks
--- (Edited on 3/27/2014 2:06 pm [GMT-0500] by Manu) ---
--- (Edited on 3/27/2014 2:54 pm [GMT-0500] by Manu) ---
Hello nsh,
Thanks for replying. 
Well simon looks to me like a dead or at least asleep project, most of the project sites are not updated since 2012. Anyway how could i reach him most effectifly ?
My next approach would be to train the german model with simon and use it with pocketphinx, if simon does not get fixed.
But here starts another question for me:
- How to add new words who are not in the dict ?
The german model (the actual LM file who says what words there are) is compiled out of huge source data. As far as i understand, this defines the phoneons order, the words so to speak.
So that pocketspinks can recognise the new words i need to update the lm file.
1. If possible i want to avoid downloading all data and building my own model file from source, then with my newly added words. If this is not possible, how do i get all the german audio archives from voxforge ? Is there a centraliesed server or something ?
2. If adation is possible, how do i adapt a lm model with my new words ?
Thanks :-D
--- (Edited on 3/27/2014 3:57 pm [GMT-0500] by Manu) ---
>Anyway how could i reach him most effectifly ?
There is a contact link on his blog: http://grasch.net/node/22
>how do i get all the german audio archives from voxforge ?
http://www.repository.voxforge1.org/downloads/de/Trunk/Audio/Main/16kHz_16bit/
--- (Edited on 3/27/2014 8:03 pm [GMT-0400] by kmaclean) ---