2

Let's say I record five friends, each individually. I label the recordings 1-5. All are the same gender and the same age-range.

Now, I have another recording. I am not allowed to hear the recording. Is it possible to identify from which friend it came?

How?

2

You cannot identify the speaker just by looking at the wave or the spectrogram. Speaker recognition is a big topic with quite a bit of controversy and you need specialized stochastic tools for it.

See Mark Liberman's recent post on this: http://languagelog.ldc.upenn.edu/nll/?p=15175.

It seems that voice biometrics have come a long way but I don't see them showing up in open source speech analytic tools any time soon.

| improve this answer | |
  • Even with a small pool of possibilities? I understand it may be hard with thousands of individuals, but for <5, I didn't expect it to be problematic. Especially if all were recorded under identical conditions, in a lab. – Teusz Oct 20 '14 at 10:01
2

I think the answer may be more mundane than you were thinking. Using the 'biometrics' tag for this question may be misleading, because it might imply to some people that the answer has to do with automated speaker recognition.

But you are just asking how you, as a person, might identify your friends via visual representations of their speech. This is definitely doable, as long as you know what to look out for, you are choosing from a small set (as you mentioned), and the recordings include enough speech to provide you with the requisite information for telling your friends apart.

I look at speech all day for a living, and I have gotten to know the voices of a few individuals pretty well. I can tell who's who based on various factors, including pitch range, intonational patterns, voice quality, and dialectal traits, all of which can be gleaned from a spectrogram. @Dominik is right about automated voice biometrics not being a widely available feature in open source speech analysis tools, but as a human I have a couple of advantages over machines--I am a native speaker of human languages and I have knowledge about phonetics and phonology--that enable me to distinguish the important parts of the signal from the unimportant noise.

| improve this answer | |
  • That makes sense. I was thinking about it in general not in this scenario which didn't seem like something they would come up that often. But how much of a sample would you need and how reliable would the guesses be? – Dominik Lukes Oct 20 '14 at 22:29
  • @DominikLukes it depends entirely on the individuals involved and how similar their voices are. The stipulation that they are all the same gender and in the same age range would of course narrow the scale of variability, but there could still be systematic cues that would be pretty reliable. Just to give a concrete example, if I know that one person has a cot-caught merger and another doesn't, and I make sure that those words are in the speech sample in similar prosodic contexts, I can tell by looking at the formant structure for the vowels in those words which speaker they were spoken by. – musicallinguist Oct 20 '14 at 23:58
  • Another example--I know that one speaker's voice gets very breathy at low pitch and volume levels and another speaker's voice gets creaky, so I make sure the speech sample includes longer sentences with early nuclear accents--like "YOU'RE not supposed to bring the dog to the vet tomorrow; I'M the one who's supposed to bring the dog to the vet tomorrow." The latter sentence would naturally bring both speakers into the bottom of their respective pitch ranges, and I could instantly tell from the spectrogram whether I was seeing breathiness or creakiness. – musicallinguist Oct 21 '14 at 0:04
  • Thanks! As a general rule, are there any generic indicators that you use or is it all speaker-dependent? – Teusz Oct 22 '14 at 7:31
  • Well, I've never actually been faced with this exact task; I'm just confident that I could do it. I would use the indicators that I mention above in my answer, but it would indeed be highly speaker-dependent and, as I indicated in my comments above, I'd have much more success if I got to choose what utterances were recorded. – musicallinguist Oct 22 '14 at 14:29

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.