Audio-Generation-using-Neutral-Net

For this project, I worked on Kaggler to execute and run my program:

Kaggle URL : https://www.kaggle.com/code/louisminguet/audio-generation-using-neural-net

According to the various tests that I was able to carry out, we observe that by simply using LSTM the results are the least interesting because many notes are repeated.
By using Embedded LSTM we get less repetitions of notes, the result is more convincing, but remains approximate.
With GAN, one obtains slightly more harmonious results than Embedded LSTM, but progress remains to be made.

Audio Generation Using :

LSTM
LSTM using Embedding
GAN

Default parameters used

LSTM :

Epoch : 70
Dropout : 0.5
Activation function : Relu (dense: 3)
Loss : Mae
Optimizer : Adam
Batch size : 256

Embedded LSTM :

128 notes
Epoch : 200
Embed size : 100
Dropout : 0.3
Activation function : Relu (dense: 1)
Optimizer : RMSProp

GAN :