A small project to test out whether is it the model or dataset causing wrong predictions.
Originally, I have recorded a audio dataset for a speaker classification project. However, it turned out bad. The chances of getting wrong prediction is relatively high. So there can be only two possible reasons：
- Bad model structure.
- Bad quatity or bad quality dataset.
After some research on another similar audio classification projects, I assume my problem is tend to be a bad dataset. So, this is small project is to test whether is it the model structure or poor quality data in the dataset casuing the miss prediction.
Basically, I won't change the model sturcture I originally desinged. But test it with a public dataset. Public datset source: https://www.kaggle.com/kongaevans/speaker-recognition-dataset
I designed a CNN model with tensorflow. Using Librosa to extract the MFCC features. And output as a json file. Split the data in the json file into train, validation, test sets for a cross validation later on. The pridiction print output includes all similarity percentage of each class. And will choose the highest percentage index.
Surprisingly, I got really good result getting all the right indexes with the original model structure but with public dataset. Conclude that it is my dataset causing the problem.