Methodology

Analyse

• Waveform

  

• FFT Spectrum

  

• Log Spectogram

  

• MFCC

  

---

Preprocess

• divided each .wav file into 5 segments to increse training data.
• exctracted mfcc for each segment.
• stored the complete dataset in format :

  {
      "mapping" : ["classical","blues",...],
      "mfcc": [[[...],[...],...,[...]],...],
      "labels": [0,2,...]
  }

---

Simple ANN

• Architecture

  

• Result

  

• Comments :

  • As we can see, the model is overfitting.

---

Simple ANN with Dropout and Kernel Regularization

• Architecture

  

• Result

  

• Comments :

  • The model is not overfitting anymore.

---

Convolutional Neural Network

• Architecture

  

• Result

  

• Comments :

  • CNN performs very well.
  • The training data had to be reshaped since CNN required 3 dimensional input.
  • Fastest

---

Recurrent Neural Network (LSTM)

• Architecture

  

• Result

  

• Comments :

  • LSTM performs very well.
  • Slowest

---

Recurrent Neural Network (GRU)

• Architecture

  

• Result

  

• Comments :

  • GRU performs very well. Better than LSTM.
  • Slowest (as slow as LSTM)