Alslamo alaikom my brother <3
Alslamo alaikom my brother <3
I tested the onnx file without baking the cmvn (when i baked it or used it with the onnx the output was blank maybe you baked it already in the onnx )
but i think the model needs more finetuning ? when i recite slowly same as al husary reciting it gets my words so good masha'a Allah
but when i recite faster it drops words or letters
if you can send the scripts of finetuning i can finetune it in my pc or something to get a better model
if you can see the finetune scripts in the https://github.com/yazinsai/offline-tarteel for the fastconformer ar model , it can help as it gets my fast recitation better ...
i know streaming model has less accuracy than offline but i think more finetuning is better ?
I'm sorry for a lot of requests ...you already made an awesome work elhamdle ALLAH <3
and if the cmvn wasn't actually by default in the onnx file or using the tlog feature ....
Alhamdulillah, I'm so glad to hear the slow recitation is working well for you!
Thank you so much for this feedback and for sharing the link to the offline-tarteel repository. You are spot on, streaming models do face extra challenges with fast pacing compared to offline models, but more fine-tuning can definitely help.
I wanted to let you know that I am already training the next version of the model which will be better, but it will take some time. I'm working on making it catch the fast recitation much better.
Regarding the CMVN and the tlog feature, I'll look into how it's interacting with the ONNX export so we can fix that blank output issue.
Barakallah feek for the support and helping make this project better!
Thanks my brother <3
i am finishing the flutter project , i'll upload it to github insha'a Allah
you can test the onnx there , also you can replace it with the new one to make a real accuracy test in your phone insha'a Allah
can you directly bake the cmvn in the onnx with default tlog features ?
sorry for alot of requests ...but i want to use this onnx instead of the offline one hahaha <3
"Our Lord, accept from us, You are the Hearer, the Knower "
you can make a script to finetune on google colab t4 gpu every 12 hour and save the checkpoint on google drive
or i can use my laptop gpu rtx 3050 insha'a Allah