Hyper-parameter tuning

[ashwinraj...@gmail.com]

Hello,

I’m using train_pnorm_fast.sh for DNN with default hyper-parameters (Vystadial english base script), training a Vystadial acoustic model and a custom domain language model, tested on a custom domain test set (test-WER: 42%).

But increasing the dataset size to ~3500 hours with same hyperparameters and setup, test-WER increases to 55%. Could you advice on what hyperparameters need tuning?

My current setup is here:

# Minimum and Maximum Language model weight

Min_lmw = 9

Max_lmw = 20     

# Number of stated for phoneme training

Pdf = 1200

# Maximum number of Gaussians used for training

Gauss = 19200

#Maximum Mutual Information beam

mmi_beam = 16

#maximum mutual information lattice beam

mmi_lat_beam=10.0

## DNN Hyper-parameters ###

number_of_hidden_layers=4

num_epochs=6

initial_learning_rate=0.002

pnorm_input_dim=2000

pnorm_output_dim=200

splice_width=4

Minibatch_size=128

Thanks,

Ashwin

[Daniel Povey]

Normally increasing the dataset size should always be helpful-- but perhaps the larger dataset was very mismatched?  That is- perhaps you were training on a different type of data than before?

In any case, the vystadial recipe is not being updated very often; the tedlium s5_r2 recipe is a better starting point.  We aren't using those nnet2-based scripts now, there are nnet3-based scripts using the "chain" model that are much better and faster too.

Dan

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.