Suggestion about hyper parameter tuning for my current data set

[Min Khant Maung Maung]

Hi all,

I would like to get suggestions about how to make hyperparameter tuning with my current dataset and my training settings.

I have the audio data duration of 150 hours which contains both the keyword and the sentence data for my mother language which is the low resource language. The audio file duration is from 2 to 10 seconds. Currently I use run.sh from wsj recipe with default parameter settings. The problem is that the hyperparameter settings from run_tdnn.sh from local/chain2 folder are written for multi-gpu training and I have only a single GPU.

So, I would like to ask you about how to tune these parameters.

1. num-jobs-initial and num-jobs-final are 2 and 8 in default settings. But because of the single GPU I have, do I need to change these both values to 1?

2. Do I also need to change max-param-change from 2 to 1?

3. Do I also need to change num-epochs 10 to 4 like any other scripts from mini librispeech recipe?

4. And please suggest to me how to improve my current training style? Should I use other recipes other than wsj recipe for my current data size which is a low amount of training data?

Thanks,

Min Khant Maung Maung

[Min Khant Maung Maung]

Any Help or suggestion?

[Jan Yenda Trmal]

I think the default parameters are just fine as a starting point 

ad the number of gpus -- change the parameter --use-gpu=true to --use-gpu=wait

it should run the N GPU jobs on a single GPU sequentially

y.

--
Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group
---
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+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/ced8644b-54ba-463d-a625-0ab13a22f67en%40googlegroups.com.

[Min Khant Maung Maung]

Thanks your suggestion sir. Please let me know which recipe should I use for my current training data (150 hours) size to get the maximum accuracy? I currently use wsj for training purposes but should I use mini librispeech or other recipes?

Min Khant 

[Jan Yenda Trmal]

I think wsj or swb -- they have ~same amount of hours. mini_librispeech is ~10 hrs but might also work fine, IIRC

y.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/4be50ea7-39f8-4304-98c7-7419bd488fa7n%40googlegroups.com.

[michael]

Please let me know are there any ticks and tricks about hyperparameter tuning in kaldi with respect to the data size (I mean audio duration) ?

Min Khant

[nshm...@gmail.com]

You'd better spend the time increasing the data size (even semi-supervised manner) and improving alignment than on hyper-parameter tuning.

[Min Khant Maung Maung]

Please could you show me the way which is the tutorial or the paper , about how to increase the data size in semi-supervised manner.

[nshm...@gmail.com]

What is your language exactly

[Min Khant Maung Maung]

Burmese, which is the official language of Myanmar

[nshm...@gmail.com]

Do you have youtube channels with or without subtitles in Burmese? You need to collect them.

[Min Khant Maung Maung]

 I can find a lot of youtube channels in Burmese. Then what can I do to increase my dataset with these youtube channels in semi-supervised manner? Could you show me the way or the example in more detail.