Full Thesis can be found here: https://aaronmkarp.com/files/KarpThesis.pdf
Named the Acoustic Counterfeit Machine, or ACM, this system is designed to hide speech from methods of mass audio surveillance, and to do so in such a way as to not arouse suspicion. The ACM is built in the following structure:
- A comparative database is created to calculate approximate matches between the spectral features of spoken words
- That database is used to generate training data to an LSTM (Long Short Term Memory) neural network, designed to match spectrally similar audio in a real-time context
- Live input is fed through the neural network and an ideal mask is calculated
- The calculate mask is played over a speaker, with a delay of <15ms between the reception of the input audio and the sonification of the matching mask