Solving Audio Captchas

Solving Audio Captchas using Machine Learning

Authors: Sampriti Panda, Duy Nguyen

Requirements

• python >= 3.6
• numpy
• scipy
• matplotlib
• python_speech_features: pip install python_speech_features
• sklearn
• keras (Not-needed, you can comment out the imports)
• pip install numba==0.48.0
• librosa, spectrum: pip install librosa spectrum

Generating Training Data

• We have provided around 50 train and 10 test cases per category, but you need to generate around 1000 train data to replicate our results.
• To generate data using our scripts, please cd into the training_data/ directory and run: ./gen_data.sh.
• You can also download pre-generated training data from: https://drive.google.com/file/d/19ypbdOiafc3Ocr9ltHIFjJI9uQXlEuJR/view?usp=sharing
• poc.py contains our original algorithm, which gives around 70% accuracy on digits and 50% on letters.
• poc2.py contains our improved algorithm, which gives around 95% accuracy.
• To run either of these implementations, modify the DIR_TRAIN and DIR_TEST directories to the necessary locations, and run python poc.py.