search

zjwang21 / mix-phoneme-bert

An unofficial PyTorch implementation of Mix-Phoneme-Bert
39 stars 7 forks source link

Accuracy of the model #2

Open rishikksh20 opened 1 year ago

rishikksh20 commented 1 year ago

Hi @zjwang21, Thanks for the implementation, but I curious to know what is the accuracy of your model implementation, in MP-BERT paper author mentioned to get 70 % accuracy of phoneme and sub-phoneme prediction, whats the accuracy you get after training a model so long. And also what dataset you used to this your model implementation?

zjwang21 commented 1 year ago

Hi @zjwang21, Thanks for the implementation, but I curious to know what is the accuracy of your model implementation, in MP-BERT paper author mentioned to get 70 % accuracy of phoneme and sub-phoneme prediction, whats the accuracy you get after training a model so long. And also what dataset you used to this your model implementation?

Phoneme and Sup-phoneme accuracy metric has been added to this repo. As for the dataset, I used our own data containing about 1b raw sentences in the field of Internet novels, which is not so diverse.

The accuracy: image image

After 40K steps, the accuracy seems coverage to about 56%. Maybe this is related to the data used. Training details will be posted here while it`s still training.

zjwang21 commented 1 year ago

Hi @zjwang21, Thanks for the implementation, but I curious to know what is the accuracy of your model implementation, in MP-BERT paper author mentioned to get 70 % accuracy of phoneme and sub-phoneme prediction, whats the accuracy you get after training a model so long. And also what dataset you used to this your model implementation?

After 100k steps, the acc reaches 60%