Models trained for 16kHz vs 8kHz

on

Models trained for 16kHz usually do not do great on 8kHz data.

When we train models on 16kHz  data and then try to decode data that is 8kHz, it is not gonna work well.

 Investigate: resampling when this problem occurs.

If you try to do this, you might get an error:

ERROR (online2-wav-nnet3-latgen-faster[5.5.920~1-b22b8]:MaybeCreateResampler():online-feature.cc:99) Sampling frequency mismatch, expected 16000, got 8000
add to mfcc.conf
--allow-upsample=true

Note: In my case above fix didn't help.

Comments

Post a Comment