search

tarekeldeeb / DeepSpeech-Quran

A TensorFlow implementation of Baidu's DeepSpeech architecture
Mozilla Public License 2.0
85 stars 17 forks source link

optimaizatiom the WER result by more fine-tunning #12

Closed smalissa closed 3 years ago

smalissa commented 3 years ago

assalam alikum @tarekeldeeb , @aibrahim- i need some help to make the thing more clear for me I have my own quranic data for female, it small (5665 records) i get (0.608 ) WER value with n-hidden = 2048 I want to obtain more enhancement of WER, so I try to change alpha, the beta value only I came to cross your work here and use your values as it used; I get (0.502) WER i need to know why you used n-hidden= 1024 and not used the default values( 2048)? what is the reason? some of clarification. thanks.

smalissa commented 3 years ago

@omerasif-itu if you can help pleaze?

tarekeldeeb commented 3 years ago

Your WER are too little, how many iterations do you train? From what I see, 1024 gives almost the same result and will consume less memory and CPU. This smaller network will be able to work with many android phones in real time. This is the main reason I chose 1024 only.

I hope it's clear now.