Recognizing Voice passwords using FFT
by Anuraag Rajasekhar
Contents
Background
The need for reliable voice recognition has become a universal goal ever since apple unveiled Siri 5 years ago. Even though initial version of voice assistant was sloppy people quickly realized the potential for a pocket personal assistant controlled mainly by your voice. With the improvement of the voice recognition technology, it's applications expanded from personal assistant to everything from voice dialing, appliance control, speech to text processing, data entry and security application to mention a few. I felt that an embedded application for voice password recognition would be a great addition to any security technology with the rise of IoT application like BLE.
The most common algorithm used by voice recognition systems is the Mel-Frequency Cepstral (MFC), but implementing this on a micro-controller would be very hard due to the complexity of the algorithm. I used the Cepstrum algorithm to process a recorded voice signal. It was much simpler than the MFC algorithm and the essence of the Cepstrum algorithm is given by this formula:
Power Cepstrum of signal $ =\left|{\mathcal {F}}^{-1}\left\{{\mbox{log}}(\left|{\mathcal {F}}\left\{f(t)\right\}\right|^{2})\right\}\right|^{2} $
The power cepstrum is a unique signature of the signal. For a security application users will be able to program voice passwords into a system, which will be used as reference signals to compare the voice password used when the user wants to unlock the system. This comparison can be done using correlation coefficient and say that if there is an 70% correlation with the reference signal, then the system will be unlocked.
The code
