SkyDocs/speaker-identification

How to test on a custom data?

Closed this issue · 4 comments

Hi, I really appreciate your work.

Here is my query, I have trained the model using Kaggle data that you have mentioned in the Read.me file, so after training, I want to test the model by passing an audio (.wav) file of "Nelson_Mandela" and see it's predictions.

I have done something like this, however model didn't predict correct. Did I test properly?

`path = ["/content/nelson.wav"]
labels = ["Nelson_Mandela", "Benjamin_Netanyau", "Magaret_Tarcher"]

try:
predict(path, labels)
except:
print("Error! Check if the file correctly passed or not!")`

Output

Speaker: 4 Predicted: 4 Sorry Benjamin_Netanyau

Hey, thanks reaching out.

The way you predicted is correct. We actually hard coded a value for 'Julia_Gillard' to debug the code. See line 436.
Its just checks if the user is Julia_Gillard then it prints 'Welcome' otherwise it will print 'Sorry' and the user_name.

And speaking of format, you dont have to pass the labels. You can keep it "unknown".

Hi, thanks for replying.

I have commented that condition and tried. Still, it is predicting incorrect. What could be the issue? Need more data for training?

nelson.mp3.zip

This is the file, which I'm using for testing (please convert it to .wav and test).

After training the model, what accuracy do you get?
Because it can be due to the accuracy error, the model is not 100% accurate. So it might be due to the edge case. You cant stick up to one wrong prediction.

Closing due to inactivity.