Help with number of tiedstates

[Rafael Oliveira]

Hi,

I'm training acoustic models for Brazilian Portuguese using a training corpus
with 166 hours of audio and a vocabulary with 65k words.

At this moment of the work, I'm investigating what values of densities and
senones give me better results, but strangely, varying the number of senones
does not significant impact on my WER, in fact the results I'm getting are
practically the same.

I have read some works that described the influence of the number of senones
in models and I know that for a database with 166 hours of audio and a
vocabulary with 65k words the number of senones should be close to 4000, but
strangely in my case there is no significant difference between the WER of the
model with 500 and 4000 senones.

Model with 500 senones: WER 20.28
Model with 4000 senones: WER 18.45

Does anyone know what could be wrong?

Thanks in advance ;)

[Nickolay V. Shmyrev]

no significant difference between the WER of the model with 500 and 4000
senones

It depends how you count the difference. 10% relative is a good improvement

Does anyone know what could be wrong?

I do not see anything wrong here

[Rafael Oliveira]

It depends how you count the difference. 10% relative is a good improvement

I agree that 10% is a good improvement, but it was not 10%, it was only 1,83%

did you use 10% just as a example?

[Nickolay V. Shmyrev]

relative

[Rafael Oliveira]

okay, got it.