We saw a password generator that detected the pattern in which you smashed the keyboard so to login in you would smash the keyboard and the generator would return a probability that you are the user. If above a certain probability then it allows the login. We wanted to do a similar concept but with the sound of your voice.

What it does

Let's you create an account with a username and sound samples, then you can try to login to see if the program recognises you by providing a new sound sample as a password.

How we built it

We used a siamese network trained on flac files of dictated text to attempt to determine if two given audio clips are of the same person's voice. A JavaScript program is used to record your sound thrice, then passed to and stored in a python server which verifies the suitability of the samples by comparing them to each other. When you want to log in, the javascript allows you to record your voice as a password, and the server then compares each of the stored samples against this password. If the network says that they're similar enough, you're authenticated.

Challenges we ran into

Installing and using tensorflow.

Accomplishments that we're proud of

This was our first time using tensorflow, and while we're not absolutely certain it's learning well (or even if it's learning what we want it to), it learned something! We managed to segment the project up nicely between us in order to distribute work effectively.

What we learned

To be honest, our neural net probably learned more than we did...

What's next for Spook Me

Training the model with a larger/more appropriate data-set to attempt to improve accuracy Familiarising ourselves with tensorflow as to ensure that our implementation is correct.

Built With

Share this project: