This model uses CNNs and dense layers to recognise the following commands:
'left', 'up', 'on', 'go', 'yes', 'off', 'no', 'down', 'stop', 'right'
Data contains 1 second clips of audio commands by multiple speakers with varying vocl features. An example of the same is visualised below (both as a time signl and as a spectogram)
This model uses librosa package for audio preprocessing and feature extraction and uses keras for building the nueral network architechture. THe model achives 89% training accuracy and 84% test accuracy. The plot of accuracy and loss is shown below
This model was run on Google Colab. Links to respective notebooks:
Training model : https://colab.research.google.com/drive/1EozlIiz2clYWkn1rWTW38UORThJeEBK3?usp=sharing
Testing: https://colab.research.google.com/drive/1MWKCGMXcIkRgp3r1NPwGtNTMAbB57TnD?usp=sharing
In order to run the pre-trained model use my_model.h5
Data, Model with weights and sample data: https://drive.google.com/drive/folders/1Cz29trw8XysGf2ERko3B_S7wAafc3eH4?usp=sharing