https://github.com/zvikinoza/masr

Mini Automatic Speech Recognition
https://github.com/zvikinoza/masr

fourier-transform keras-tensorflow librosa sound-processing speech-recognition speech-to-text

Last synced: over 1 year ago
JSON representation

Mini Automatic Speech Recognition

• Host: GitHub
• URL: https://github.com/zvikinoza/masr
• Owner: zvikinoza
• Created: 2020-01-04T12:45:50.000Z (over 6 years ago)
• Default Branch: master
• Last Pushed: 2022-09-26T19:29:50.000Z (almost 4 years ago)
• Last Synced: 2025-01-26T15:16:56.561Z (over 1 year ago)
• Topics: fourier-transform, keras-tensorflow, librosa, sound-processing, speech-recognition, speech-to-text
• Language: Jupyter Notebook
• Size: 84.1 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Mini Automatic Speech Recognition 

## Using Fourier transforms and CNN (VGG like) architecture

![ ](https://github.com/zvikiNozadze/MASR/blob/master/imgs/number_spectogram.png)

EDA and modeling on ≈150 samples of sounds speaking numbers form 1 to 5 recorded by 10 people.





Accuracy on large (>10k) dataset: 94%. 


Accuracy on given small  dataset: 93% 




Task was more or less chalenging because of small dataset.  




Augmentation techniques:

*   increase/decrease pitch

*   increase/decrease speed

*   stretching

*   frequency and time masking

*   white noice injection

*   time shifting

*   overlay 2 samples (quiet and louder)

*   pre/post noise padding  




Splitting data is done by speakers 5-5.







Training graph :




![ ]( https://github.com/zvikiNozadze/MASR/blob/master/imgs/acc_graph.png)