always predict "classical" in real world record wav

[Zepan]

Hi, I just test your model with wav that record by real world mic, and convert to 16k samplerate wav.
I record the song in the dataset, and env is very quiet, but what ever genres music I record, your model always predict "classical"
It seems the model is overfit too much.

[Hguimaraes]

Hello, @Zepan!

Unfortunately, I have not been active in this project for a while. I suppose that this could be related to the sample rate. I just accepted a very old pull request that I think it could fix.

If this is not the problem, I don't have a quick explanation. This is a very simple model (Proof-of-Concept), I've never tested in a real-world scenario before. If you have another dataset, with different channel conditions and noises, you probably can train a better model than this one.