https://github.com/j77m/stuffy-nose-recognition

Trained neural network for recognizing speaking with stuffy nose.
https://github.com/j77m/stuffy-nose-recognition

Last synced: 28 days ago
JSON representation

Trained neural network for recognizing speaking with stuffy nose.

• Host: GitHub
• URL: https://github.com/j77m/stuffy-nose-recognition
• Owner: J77M
• License: mit
• Created: 2018-11-17T19:05:44.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2018-11-21T18:36:27.000Z (about 6 years ago)
• Last Synced: 2024-12-14T00:29:28.557Z (about 1 month ago)
• Language: Python
• Size: 10.1 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# stuffy-nose-recognition

Trained neural network for recognizing speaking with stuffy nose. 


Extension : real time recognition on 2 seconds frames.


Model was trained only on data from one person (me), so there is possibility to record data and train model. 


Also record new data and train model if demand is smaller frame (more real time recognition).


(There was extension in training process, instead of training on recording data, to train on analysed wav files, but results on real time were about 60% accuracy, so feel free to try and improve)





|


|__ training 


|          |_ data		# data recorded by data_record


|          |_ model_evaluate	# folder for tensorboard evaluations


|          |_ trained_models	# trained models


|          |_ data_record.py	# record and saves


|          |_ model.py		# neural network model 


|          |_ utils.py		# useful functions


|          |_ audio.py		# class for audio processing


|


|__ real_time_analysis.py	# real time use of trained neural network 


	   







Testing showed, that the best architecture is :


_________________________________________________________________


Layer (type)                 Output Shape              Param #   


=================================================================


conv1d (Conv1D)              (None, 991, 64)           704       


_________________________________________________________________


max_pooling1d (MaxPooling1D) (None, 247, 64)           0         


_________________________________________________________________


conv1d_1 (Conv1D)            (None, 238, 64)           41024     


_________________________________________________________________


max_pooling1d_1 (MaxPooling1 (None, 59, 64)            0         


_________________________________________________________________


flatten (Flatten)            (None, 3776)              0         


_________________________________________________________________


dense (Dense)                (None, 256)               966912    


_________________________________________________________________


dense_1 (Dense)              (None, 1)                 257       


=================================================================


Total params: 1,008,897


Trainable params: 1,008,897


Non-trainable params: 0


_________________________________________________________________


None





batch_size = 5, epochs=10 


100% validation accuracy on my data (i wont share them, because of my privacy :) )


data - total: 4.5 min of recording (2.25 min talking with stuffy nose - nose stuffed with fingers, 2.25 min talking with clear nose)





If you are training model, please test and evaluate your combinations of architecture.


Training code is adapted for tensorboard evaluation. 





J.M.