noise augmentation scripts

[Sravani Lekkala]

I have trained on 50 hours of data.  It is giving good results for clean recordings. But it is giving some random results for the background noise when the user is even silent.  can anyone suggest the method to solve this issue?   Shall i add noise to 50 hours of data and combine all the recordings then train from starting? Where can i find the scripts for noise augmentation?

[Raj Gohil]

Hello,

You can augment the data with publicly available MUSAN data on openslr.
and augmentation scripts can be found in chime6 run.sh in the earlier stages.

Thank you
Raj Gohil

[Sangramsing]

I am little surprise why are not creating anyone videos as a step-by-step process where download the data where change exactly and how to deploy our data and physical operating system and which file it’s needed to modify and then run khadi recipe

--
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/7ec4bc94-8761-46c8-a98c-43db86b11a05n%40googlegroups.com.

--

Sent from my iPhone
Thanks with regards,

[Sravani Lekkala]

Will it work for any kind of background environment?

--

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 a topic in the Google Groups "kaldi-help" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/kaldi-help/iMW6X5iX9sM/unsubscribe.
To unsubscribe from this group and all its topics, send an email to kaldi-help+...@googlegroups.com.

[aliiire...@gmail.com]

Hey.

Here an example f augmentation data in Kaldi, please read the script and follow it step by step for your dataset.

https://github.com/kaldi-asr/kaldi/blob/master/egs/swbd/s5c/local/nnet3/multi_condition/run_aug_common.sh

augmentation of data usually doesn't change results in in-domain test dataset but makes model robust with environmental conditions, noise, reverb, ... .

If your dataset small, like 50 hours you can use 3-way speed perturbed plus augmentation.

"any kind of background environment" is too common and cant confirm it all condition but it is Ok for most environments.

best regards