search

MiteshPuthran / Speech-Emotion-Analyzer

The neural network model is capable of detecting five different male/female emotions from audio speeches. (Deep Learning, NLP, Python)
MIT License
1.3k stars 438 forks source link

Accuracy problem #27

Closed torrentillo closed 5 years ago

torrentillo commented 5 years ago

Hi Mitesh, I’m trying to obtain the 70% accuracy you got but I’m only getting a 35% could you please tell me the database and send me the exact code you used to get the 70%?? Thank you very much. My email is listor69@gmail.com

MiteshPuthran commented 5 years ago

Hello @torrentillo, this is the exact code that I had used to train my model. In order to achieve higher accuracy, there are 3 things that come to my mind.

  1. Try playing with the parameters while building the model and test it out.
  2. Extract more features from the audio files. This can be done by increasing the sampling rate.
  3. Add more data by copying the existing files. (Not the best way, but can improve the accuracy)

Hope this helps.

torrentillo commented 5 years ago

Hi Mites,

I used the exact code and I can't get the accuracy you got, what about the dataset you used? could you please update the dataset you used to this post??

MiteshPuthran commented 5 years ago

I had used the dataset from the links I posted in the README file.

torrentillo commented 5 years ago

Yes, but when you filter only by 5 emotions you have like 1300 audios for training instead of 900 that I have if I do the same, so you've trained your model with some extra data that is not in the like you posted in the README file.

Thank you

beachboysqq commented 5 years ago

Hi,I am trying using different models and increase the dropout value.But it seems still overfit. I can get a hign acc on training but low acc on test. Plz give me some suggestions~ many thanks~

torrentillo commented 5 years ago

Hi beach,

I've achieved a 70 % accuracy with this model :

Ridge = sklearn.linear_model.Ridge()

nikhil2593 commented 5 years ago

same..not getting specified accuracy with the provided model architecture and dataset.