2

I used an automated customer service system that requires reading off one's case id. The case id I had included the sequence "3F."

The speech recognition software was only tripped up by that "F," which it asked to clarify if it were an "S" or an "F." I am assuming it picks the n-most likely phones and asks the user to choose when it meets some threshold of uncertainty.

When I spoke to an actual human (on the phone), she also asked if it were "S" or "F."

My conclusion is that having a word which begins with a interdental voiceless fricative causes a following labiodental fricative to be confused. There were other "F"'s in the id, but these were not confused (nor did they come before any interdental sounds).

Improve this question
2
  • 2
    No, it's just that the difference between /θriyɛf/ and /θriyɛs/ is only perceptible at the end of the word, and it's very hard to distinguish because of the fricative bar blanking out the high frequencies. This would only occur when naming letters, which is a very low-tech way to communicate. If you must use letter names, try using names that are easily distinguishable; that's what the military "Alpha, Bravo, Charlie, Delta" stuff is about.
    – jlawler
    Nov 28, 2014 at 19:10
  • Yes basically the acoustics of "eff" and "ess" are similar and if they're using context sensitive software the context "following 'three'" doesn't help disambiguate because neither pair is more common or relevant.
    – hippietrail
    Nov 29, 2014 at 12:37

1 Answer 1

Reset to default
6

As @jlawler mentioned in his comment, word-final [f] and [s] are likely to be confused over the phone in general. The place of articulation of a fricative is mainly cued by two things--(1) the center of gravity (or diffuseness) of its noise and (2) the direction of the formant transitions going into and coming out of it.

In the case of word-final fricatives, if there is a pause or a glottal stop after the word, there are no formant transitions coming out of the fricative, and the transitions going into it are often obliterated by an early cessation of voicing in the vowel that helps cue the voicing of the fricative. So often the strongest cue that distinguishes a word-final [s] from a word-final [f] is the fact that the [s] is realized with high-intensity noise in the high-frequency end of the spectrum. However, the filter imposed by the phone attenuates the noise in that high-frequency range, rendering [s] noise and [f] noise much less distinguishable.

As for why just that one F in 3F was confused, it is highly unlikely that it had anything to do with the interdental fricative at the beginning of 3. Consonant perception (both by humans and by machines) doesn't work long-distance like that. Either it was a coincidence and you happened to provide weaker cues on that particular [f] coda, or there was something different about the context after the fricative that rendered it more ambiguous.

Improve this answer
3
  • As a follow-up, are machine phoneme perception errors transitive, e.g. if a machine has difficulty with some sounds, will it always be the case that a human would too (over a noisy environment like the phone)? Or are there some perception errors that would be expected for a machine but not a human, i.e. the set of sounds which are ambiguous is not the same for both machine/human?
    – user3898238
    Nov 28, 2014 at 23:38
  • +1. I first discovered this problem when I had college friends named 'Jeff' and 'Jess'. Phone conversations among their mutual friends were frequently derailed by this problem. It wasn't even that we couldn't tell which one was mentioned -- what made it such a problem is that we would often think we'd heard one name when in fact the other had been said.
    – ruakh
    Nov 29, 2014 at 0:16
  • @user3898238 I'm more of an expert on human speech perception than on automated speech recognition. I can tell you with confidence that the types of errors an automated system makes depends on the algorithms used by the system. I'm also going to go out on a limb and say that the sets of sounds that are ambiguous over the phone for machines and for humans are probably quite similar in general, but that humans have a leg up because they can overcome some ambiguities with certain types of pragmatic expectations that aren't programmed into the machines.
    – musicallinguist
    Nov 29, 2014 at 4:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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