Me and my friends are working on a personal guardian project that can be activated using voice and used CNN and some other methods but our model is over-fitting. any suggestions?
It is nice project. I would like to suggest you some recent work related to your project based on sound data: https://medium.com/@mikesmales/sound-classification-using-deep-learning-8bc2aa1990b7
Article Detection of Pathological Voice Using Cepstrum Vectors: A De...
www.mdpi.com › pdf
Usually the best features for speech and voice recognition have been over the years found to be related to Mel Frequency Cepstrum Coefficients (MFCC) since the Mel scale has been defined keeping in mind human hearing and auditory perceptions. To avoid over-fitting, one method frequently used is cross-validation i.e. splitting into multiple train-test sub-datasets and tuning the classifier parameters to reduce classification error.
This chapter can help:
Chapter Voice Identification Using Classification Algorithms
You can use Hidden Markov Model (HMM)
Article A Tutorial on Hidden Markov Models and Selected Applications...
I wish to suggest you carry-out your experiment using Recurrent Neural Networks (precisely GRU) as this category of Deep Learning models were primarily designed for speech recognition.
Bonaventure Molokwu can you suggest me a good paper or book on GRU?
- Badges
- Science topic