Prediction failure

[Mlallena]

Hello.

I am testing your idea, and am trying to create a network that can identify between three different languages. However, I am encountering a few problems.

The first is that the precision of the model never seems to rise above 65%. Could this be a matter of not having enough data to work with?

The second is that, when I try to use one of the saved states to make a test (using the same procedure you mention here) the prediction result is always the same (0.380580,0.269690,0.349730), no matter what spectrogram I use. What's the issue here? How can the system return the same prediction for very different audios?