0
$\begingroup$

Do any general-purpose techniques exist to distinguish live spoken speech from played-back speech samples? For example, if I were to deploy a speaker verification system to authenticate users by voice, I would not want the system to be able to be spoofed by playback of voice recordings.

I think I can imagine at least one technique which might work for certain known combinations of microphones / speakers: If a microphone of sufficient quality captures a novel signal outside the speech band that corresponds (in time) to a signal within the speech band, this would seem to indicate a high likelihood of the sample not being a live speaker (because live speakers don't emit signals outside the speech band). Is this reasoning sound? (No pun intended).

$\endgroup$
3
  • $\begingroup$ and where would that out-of-band signal come from? $\endgroup$ Aug 30 '19 at 19:52
  • $\begingroup$ E.g. playback via CRT television? Could be anything; I don't think it's particularly important $\endgroup$
    – Descartes
    Sep 3 '19 at 13:21
  • $\begingroup$ well it makes all the world of a difference: as far as the math is concerned, playback can perfectly reproduce a band-limited signal such as speech. You really need to come down with very specific definition of your signals! $\endgroup$ Sep 3 '19 at 14:42
0
$\begingroup$

Do any general-purpose techniques exist to distinguish live spoken speech from played-back speech samples?

No, at least not how I understand your definition.

Consider the case of "live speech" : some one talks into the microphone, you digitize the signal and analyze it. You can also store the digitized speech and analyze it a day later. Now the exact same samples constitute "play back". Since the samples are identical, you can't distinguish them from "live speech".

$\endgroup$
0
0
$\begingroup$

You have to time stamp the speech with something that did not exist in the past. For instance, generate a cryptographic random number, and compare the spoken response latency to the speed of light over the distance to speaker.

$\endgroup$
2
  • $\begingroup$ Not saying that this also can’t also be spoofed. But you are now entering the realm of adversarial DNNs, where the one with biggest training set and largest TPU cluster wins. $\endgroup$
    – hotpaw2
    Aug 31 '19 at 4:26
  • $\begingroup$ Interesting idea. I'm not sure I understand what you're suggesting though. Could you elaborate? What would the random number be used for? Wouldn't I want to "compare" to the speed of sound, not light? How would one obtain the spoken latency? $\endgroup$
    – Descartes
    Sep 3 '19 at 13:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.