(Apologies in advance for the expected misuse of terminology; I am not a linguist. Please correct as appropriate.)

I am considering a Deep Learning language evolution experiment and would like to include in the model consideration of the difficulty of producing particular sounds, singly and in pairs, possibly triples. So that, for example, /b/ /g/ /t/ would be assessed as harder to produce (hence probably lower communication rate for such combinations) than /b/ /a/ /t/.

The question is: where can such information be obtained, or how could it be created from other existing resources?

I believe I should prefer consideration of phones over phonemes because any one phoneme may have multiple allophones, however, if only phoneme metrics are available I shan't complain.

It would be interesting to consider phones in all generality, and to allow for the possibility of mixing e.g. Xhosa glottal stops with tonal (tonemes?) characteristics, but that might be a bit ambitious initially, so metrics for specific languages or language groups would be perfectly good to begin with.

(At this stage of the modelling concept, I also doubt that I shall be considering other things like modification of sounds according to their context.)

1 Answer 1


Segments are not intrinsically hard, or easy, to produce, so there is no ranking of segments. It may be possible to derive a subject pool-specific ranking, where a pool of subjects have a difficulties with {a,b,c} but not with {d,e,f}, and the content of those sets is related to the languages that they speak. Difficulties may be perceptual or they may be articulatory, or both, and the two causes are often related.

Certain people have difficulty with [θ,ð], though generally not English, Greek or Icelandic speakers. You can detect this by observing non-fluent productions of English where "this" may be produced as [zis] or [dis]. There is also a common problem with [r] and [l], a speaker of another language may mistake "lip" and "rip". Perception difficulties are more common, and are often the cause of articulatory errors. Because of this chicken-and-egg question, it is generally pointless to focus just on articulatory mechanisms.

There has long been a belief (not totally crazy) that there is an ordering to segments, whereby that t is more basic, indispensible, easier or however you want to put it, that d and d is better than t'. This notion is embodied in the theory of what is known as "markedness". There is no clearly agreed-on ranking of segments (consonants and vowels are generally treated as completely unrelated), but one might for example be able to order the IPA to reflect a hierarchy like t>k>p>q>ʈ>ʢ, generally based on frequency of occurrence across languages. The general (and extremely wide-ranging) topic of segmental markedness will most likely yield a context-free ordering of segments, but applicability of such a ranking will face language-specific quirks such as the lack of /p/ in Arabic and the presence of rather "marked" /tˁ ʕ ħ/.

The example that you provided of /b/ /g/ /t/ being harder that /b/ /a/ /t/ indicates yet another problem, that you are interested in certain sequences of segments and not just the segments /b/ /g/ /t/ is a list of three sounds (the order doesn't matter), and /b/ /a/ /t/ is another list of three sounds – mostly overlapping the first set. Therefore what you must mean is that you think the sequence /bat/ is easier than the sequence /bgt/. This, fortunately, is also reasonably well studies, though is generally about perception and not about production. Phonologists often organize sequences of segments into syllables, which follow rules, and as a general rule, a syllable is centered around a vowel, with some preceding and perhaps following consonant(s). /bat/ can be syllabified as is in English, but not in Logooli (nothing after the vowel in the syllable). English allows /brat/ but many languages (which have all of the segments) don't allow initial clusters (Logoori is an example). Reading up on syllable structure and the notion of "sonority sequencing" will reveal some of the main generalizations (and the controversies).

Articulation can sometimes be brought to bear and given a bit more advantage in constructing an explanatory account for a fact. There is no difficulty in producing the set of constrictions in /bgt/, but there is an aerodynamic-physiological challenge in making /g/ overlap /b/ so that there is no release of air in the transition from /b/ to /g/. This is compounded by the need to keep pumping air out of the lungs to produce the span /bg/. We don't do this in English so we don't have articulatory practice with such a sequence, thus it is most likely that this would be produced as [bɨ̆gt].

If you aren't considering how segments are modified according to context, then it's very non-obvious that you can model anything validly. Absolute markedness is pretty unimportant in language change, it's all about context.

Xhosa doesn't have glottal stops, so that will simplify things, unless you mean something besides glottal stop.

  • 1
    Thanks; let me digest that; but in the meantime, yes I meant sequences, no I didn't mean glottal stops I meant clicks (but I was thinking of a glottalised click at the time), yes I was thinking of the articulation difficulty - two successive sounds that differed only in the placement of the tip of the tongue seemed likely to be easier/faster to produce than if the whole tongue had to move from forward & up to back and down. But, if I understand correctly you are saying: everything is equally easy with practice. Commented Jan 22, 2021 at 16:44
  • It's not so much ease as it is habit. English phonotactics (permitted, required, and forbidden combinations of speech sounds) doesn't allow words to start with /fn/, though /fl/ is OK. Other languages have no trouble with /fn/. It's the set of habitual speech rules, and it's not independent of -- but not dependent on, either -- the difficulty of pronunciation.
    – jlawler
    Commented Jan 23, 2021 at 0:46
  • @jlawler "It's not so much ease as habit" Can you clarify what "it" is?. Crudely, for machine learning one has costs (-) and benefits (+) - and solutions with net +ve are to be preferred; it is very important for me to know either that all sound sequences are equally easy - or to quantify the differences. "Habits" may emerge but "ease" should be ~physiological constraint Commented Jan 23, 2021 at 12:13

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.