I would like to verify that what I am doing is right, although training the acoustic model just doesn't finish and gets discarded.
I am working on pronunciation variations for the same phone. For instance, Q can be pronounced differently such as K and G by some dialects. What I did was I created a pronunciation dictionary that now contains variations of the sound wherever it occurs in the transcription. I then added these new phones to the phones list. I then ran pocketsphinx using default paramters 8 for gaussians and 200 for senones (will modify them later, I just want to make sure that the training goes through and I get the tested data decoded). In Module 20 Training Context Independent Models, pahse 3: Forward-Backword, my training goes 1 Gaussian, iteration (1) through 1 Gaussian, iteration (6), it then goes to Module 30 Training Context Dependent Models and in phase 3 it does only Baum Welch iteration 1, then it says Training failed in iteration 1. However, it doesn't continue to 2 Gaussians, then 4 Gaussians, then 8 Gaussians.
Since my phones list contains 48 sounds including the newly introduced ones, what could be the problem and why do you think training doesn't go until the end.
Best Regards,
Abushariah, Mohammad
Last edit: Mohammad 2017-03-26
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Dear All,
I would like to verify that what I am doing is right, although training the acoustic model just doesn't finish and gets discarded.
I am working on pronunciation variations for the same phone. For instance, Q can be pronounced differently such as K and G by some dialects. What I did was I created a pronunciation dictionary that now contains variations of the sound wherever it occurs in the transcription. I then added these new phones to the phones list. I then ran pocketsphinx using default paramters 8 for gaussians and 200 for senones (will modify them later, I just want to make sure that the training goes through and I get the tested data decoded). In Module 20 Training Context Independent Models, pahse 3: Forward-Backword, my training goes 1 Gaussian, iteration (1) through 1 Gaussian, iteration (6), it then goes to Module 30 Training Context Dependent Models and in phase 3 it does only Baum Welch iteration 1, then it says Training failed in iteration 1. However, it doesn't continue to 2 Gaussians, then 4 Gaussians, then 8 Gaussians.
Since my phones list contains 48 sounds including the newly introduced ones, what could be the problem and why do you think training doesn't go until the end.
Best Regards,
Abushariah, Mohammad
Last edit: Mohammad 2017-03-26
Instead of asking us to divinate with coffee beans you should have provided the data to reproduce your problems.
do you work on Arabic language ? if so, why are you turning the phonemes to english?
It is recommended in tutorial.
Ah ok ... I'm saying this because I'm working with Arabic phonemes and it worked well for me .. I don't know why it's recomended to change it
It might work for you but overall it is helpful to have english names for phonemes.
HI
I think rawan is right , you can work with every thing in arabic however you should take care of unicode issue if you are using windows