V.0.2.0 streaming no more standarizing MFCC, is it better?

[mar_martinez]

Hi,

With the in-progress streaming native_client v.0.2.0 I observe that MFCC in audioToInputVector are not standarized anymore, neither for training nor for inference (of course), I assume that this is because the streaming make it more difficult because a standarization window should be set to the the average and std figures.

So I think that the trained models I had are not compatible with this version.

But the new models I’ve trained perform worse than the older ones, using the same data and tuning hyperparameters.

Could it be because of the standarization? Did you get the same WER with this version in your benchmarks?.

Best regards,
Mar

[reuben (Reuben Morais)]

I assume that this is because the streaming make it more difficult because a standarization window should be set to the the average and std figures.

Yes.

Did you get the same WER with this version in your benchmarks?

We have achieved the same WER without normalization, but due to changes to the language model it has since gone up.