Hguimaraes
commented
2 years ago 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.
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.