Voice Authentication with FFT

Status
Not open for further replies.
M

mala96

Member
Hello dear Forum members,

i have an idea for an algorithm i want to realize at an uni-project. And i would like to know what you guys think about it before starting.

I've seen a thread regarding FFT support for teensy 4.0 in this forum. Paul Stoffregen posted a segment of code there which should give the FFT spectrum of an audio signal which is beeing sent through PIN 2 of Teensy 4.0 or Teensy 4.1.
So that made me think ... could there be a way to connect a microphone to PIN 2 and record some audio as a fourier transformation and kind of save it as a sample ... and then make another recording and start to compare the discrete frequencies with the first recording?

With an algorithm like that you should be able to do some kind of voice authentification and i would like to give that project a try.

does someone know the FFT code origin from Examples>Audio>FFT? I would need to sweep through the spectrum and get a matrix of discrete values for specific discrete frequencies, rather than a graphical result for a spectrum analyser.

Furthermore: has someone an idea on how to evaluate the data recorded in recording 1 and recording 2?

Any help or suggestions about this would be highly appreciated :)


with kind regards,
Mala
 
Hi,

not sure where the "Pin2" comes from. Anyway, It's not important.
I believe that it is possible to distinguish different voices. At least maybe women's and men's voices?


I would take a teensy 4.1 with an audio board, along with the microphone.

You should look at the audio library and examples for that though.
You don't need to touch or change the code for the FFT. If you are interested, you can find it here and here.

Interesting project!

I'd try it without a microcontroller at first. Maybe use audacity and matlab?
 
Frank B said:
Hi,

I believe that it is possible to distinguish different voices. At least maybe women's and men's voices?
Click to expand...


Guten Morgen Frank,

i thought of using this device for example for my phone as an authentification.
With a FFT of my voice I should get a pretty unique fundamental with a overtone series. I think it could work pretty well.

Audacity and Matlab might me a good start indeed! Maybe I'll sample a lot of me talking with an old MIC and analyse how the spectrum is behaving when I'm talking. Then I would just need to write some kind of sorting algorithm that compares the original sample with the sample that "wants to be authentificated" and the device would be finished :)

Found some interesting projects on research gate aswell:
if some of you are interested google for 'Voice Authentification site:researchgate.net' there are lots of good approaches.


One thing still is bothering me: what if someone has an audio record of my voice? how could i avoid that person hijacking my phone when they play that record with another device? Do you guys know if a microphone can sense the ankle a person is standing in front of the MIC? That way i could at least ensure that a playback without a real person talking wouldn't be mistaken as my real voice.

Thank you all!
 
mala96 said:
One thing still is bothering me: what if someone has an audio record of my voice? how could i avoid that person hijacking my phone when they play that record with another device? Do you guys know if a microphone can sense the ankle a person is standing in front of the MIC? That way i could at least ensure that a playback without a real person talking wouldn't be mistaken as my real voice.
Click to expand...

As one possibility for avoiding the problem of distinguishing between "live or Memorex" (showing my age), you could have the Teensy randomly generate & display a combination of a few words from a much larger list to be spoken whenever you need to authenticate. You could "train" the system in advance with your voice for each word & require that the potential speaker speak these words in the given order, then check for a quality match. Someone else would not only have to know your entire list, but to also have a recording of you saying each & every one of them in order to completely fool the system. Just a thought . . .

Mark J Culross
KD5RXT

P.S. Another thought with a simple example of making it even more foolproof: if you used colors as a simple list of words, when training the system, you could intentionally speak a color that is different from the one that is displayed, making it even harder for someone else to pass your test. You just have to remember the mapping from displayed color to spoken color. MJC
 
Status
Not open for further replies.
Back
Top