Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Low accuracy value #2

Open
adamchaabani opened this issue Aug 15, 2022 · 6 comments
Open

Low accuracy value #2

adamchaabani opened this issue Aug 15, 2022 · 6 comments

Comments

@adamchaabani
Copy link

Hi, I am running the stress detection code.
First, I can't apply the alignment on the database after removing the stopwords there is an error and second and without removing the stopwords and after the train of the model I had an accuracy on the train 37% and on the test 36% which is very low compared to your project documentation, normally 89%..

@vivekkr12
Copy link
Collaborator

vivekkr12 commented Aug 15, 2022

Since this model is essentially a binary classifier - detecting whether a particular phoneme has primary stress or not, the baseline accuracy should be ~50% without any training.

Did you follow the steps mentioned in README on how to train the model? Refer to this issues on some hints on re training the model. #1

What error are you getting while running alignment?

@adamchaabani
Copy link
Author

Capture d’écran du 2022-08-15 19-36-57

here is the error i got when i removed the stopwords and applied the alignment

@adamchaabani
Copy link
Author

Yes sir I followed the readme file step by step except that I only worked with the librispeech database because I could not find where to download the OSCAAR database.
After the train test split stage I deleted the secondary stress file (2) and applied the train by varying the hyperparameters and still low accuracy. Could you help me please?

@vivekkr12
Copy link
Collaborator

Phoneme alignment doesn't need any stop words removal, all you need is a dictionary file. We skip feature generation of the stop words.

@adamchaabani
Copy link
Author

the problem is that even without removing the stopwords it gave very low accuracy results compared to your project documentation, normally 89%. I don't know where the problem comes from

@vivekkr12
Copy link
Collaborator

What accuracy do you get while evaluating on test dataset without training? It should be ~50%.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants