Chinese models trained on a huge amount of acoustic data

[yeelearn]

I want to know the data set used by the Chinese(mandarin) model provided by CMU.

When I want to train my Chinese model, I find that the recognition rate is very low.
I suspect that the data set is of poor quality or training method problem.

[Nickolay V. Shmyrev]

I want to know the data set used by the Chinese(mandarin) model provided by CMU.

https://catalog.ldc.upenn.edu/LDC98S73

When I want to train my Chinese model, I find that the recognition rate is very low.
I suspect that the data set is of poor quality or training method problem.

If you use AISHELL http://www.openslr.org/33/ or AISHELL2, the accuracy should be much better.

[yeelearn]

thanks

I want to know what your noise file is, such as +LAUGH+ +LIPSMACK+ +COUGH+ ...

would you mind sharing “sphinx_train.cfg” file with me, so that I can find different places with my files.

[Nickolay V. Shmyrev]

I want to know what your noise file is, such as +LAUGH+ +LIPSMACK+ +COUGH+ ...

Noisedict inside model contains:

<s>         SIL
</s>        SIL
<sil>       SIL
++laugh++   +LAUGH+
++lipsmack++    +LIPSMACK+
++cough++   +COUGH+
++breath++  +BREATHE+
++incomplete++  +GARBAGE+

would you mind sharing “sphinx_train.cfg” file with me, so that I can find different places with my files.

The model was trained long time ago, so the configuration file is lost. You can use default one, it should give you same results.

[yeelearn]

Ok, let me try again.
Thank you very much.