How do linguists place the vowels of a language precisely on the vowel trapezoid?

Question

Since vowels in human speech are a continuous spectrum rather than a discrete set, many descriptions of languages I’ve seen — not only on Wikipedia — place the vowels of a language as dots in a two-dimensional grid. For example, here is the one for modern Hebrew:

The positions of these dots seem to be pretty precise; in charts for other languages, dots marked with the same IPA characters are nonetheless in slightly different locations. For example, compare the locations of [a] and [i] in the above to those in the French one:

However, as we know, everyone learns as a child to categorise vowels they hear into the phonemes of their native language. This makes it extremely hard for any particular person to hear subtle distinctions such as the one between the Hebrew [i] and the French [i]. It would seem likely that two linguists of different native languages would place the above dots in (slightly) different places when analysing Hebrew.

So how do we actually know where exactly to place these dots? Is there some form of modern technology that is used to analyse vowels? Is there perhaps a software that can be downloaded which would allow the user to speak a vowel into the microphone and have it rendered on such a diagram?

Answer 1

There are, in fact, some rough and ready acoustic measures for analyzing vowels. We call them formants, and their acoustic theory is a little heavy to get into right here. Roughly, they are bundles of re-enforced harmonics: Formant (Wikipedia). It's harmonics that make the same note sound different when played on a violin, a viola, and a piano, and it's harmonics that make /i/ sound different from /u/.

The first two formants, abbreviated F1 and F2, correspond roughly to vowel height (F1) and vowel backness/rounding (F2). Automatic formant estimation from raw audio is a tricky business, but it's a project that I myself have been involved in.

So, the way I analyse vowels now is

1. Obtain a relatively long speech sample from a speaker.
2. Transcribe the speech sample.
3. Run the speech sample + my transcription through our software, which produces estimates for F1 and F2.
4. Attempt to normalize the acoustic measurements to account for differences caused by different head and body sizes. Anatomical differences like these cause big differences in the raw measurements.
5. Make a plot.

Now, that's probably a little different from how phonological descriptions like the one you posted work. I'd imagine most of these plots with apparently precisely placed points are based partially on acoustic data like I just described, and partially on the phonologist's view of the systematic character of the sound system. I say that because plots of acoustic data are rarely ever so clean as that plot, with perfectly parallel /i/ and /u/, /e/ and /o/.

Answer 2

[I thought I'd add this answer to cover in a more systematic way some of the points brought up in earlier answers]

The first point to make is that trained phoneticians are able to hear the differences between similar-sounding vowels in different languages, although these are not noticed by casual observers.

Concerning how to make vowel plots, there are, in principle, several ways that a vowel trapezoid could be constructed:

1. Based on objectively measured articulatory characteristics of the vowels
2. Based on proprioceptively assessed articulatory characteristics of the vowels
3. Based on auditory characteristics of the vowels
4. Based on acoustic characteristics of the vowels

Examples of all four types of vowel plots can be found in Lindau (1978). The first method has been used in the past with varying methods which would be considered too invasive by today's standards (x-ray motion pictures, placing gold or lead chains, buttons on the tongue, etc.). Modern techniques such as ultrasound or MRI are still used in well-equipped phonetics laboratories, but on whole method (1) is not practical for ordinary descriptions.

Method (2), called the "Cardinal Vowel" method, is a bit curious in historical perspective: it requires a phonetician trained in the method to repeat a vowel sound uttered by a native consultant until the phonetician's imitation is nearly perfect, and then to proprioceptively assess his/her tongue position in articulating the vowel (Abercrombie, 1991: 40). What is interesting about this method is that in order to use it you would have to be personally trained by the late Daniel Jones or by one of his students, and this is why it has more or less died out. But to get an idea of how it went, consider a typical description of some vowels in Tswana written by Daniel Jones:

The Sechuana o is very nearly cardinal vowel No. 7 (The French sound of ô in tôt); the tongue-position is if anything, a shade lower than this. The sound does not exist in Southern English, but may be heard in the Scottish pronunciation of words like home, go. It must be carefully distinguished from the diphthongs heard in the various English pronunciations of such words. (Jones and Plaatje, 1916: xxiv)

The third method is rarely done in a strictly controlled way because of the difficulty in measuring auditory properties precisely. It is, however, used informally by phoneticians with a trained ear. Ladefoged has claimed that many phoneticians claiming to use the Cardinal Vowel method were actually using a sort of auditory based method (most likely because it is actually difficult for anyone to feel the position of their tongue body):

. . . after you’ve used a term like tongue height—raise the tongue, lower the tongue, move it to the back or front—you begin to feel that that is what your tongue is actually doing. But you are kidding yourself. This is not really what one does in trying to produce a vowel. (Fromkin, 1985: 5)

This leaves, finally, the vowel plot based on acoustic characteristics of vowels. Most vowel plots you will see published today are constructed by measuring the first and second formant values of vowels, and either using these values directly or using some kind of transformation of them (e.g., a derived perceptual scale or a normalized value) in making the plots. The main difference between the vowel plot and the kinds of plots you make in high school algebra is that the axis originates at the upper-right corner. Acoustic "backness" is a function of F2 (or F2-F1), and moving to the left corresponds to increasing F2. Acoustic "height" is a function of F1, and moving down corresponds to increasing F1.

Software exists for tracking formants, but it is not completely reliable and its results need to be double-checked by a trained phonetician. Fortunately, it is much easier to learn approximate formant values for different vowels than it is to learn how to use the Cardinal Vowel method.

David Abercrombie (1991). "Daniel Jones's teaching." Fifty Years in Phonetics: Selected Papers. Edinburgh University Press. pages 37–47.

Victoria A Fromkin. Introduction. In Victoria A Fromkin, editor, Phonetic Linguistics: Essays in honor of Peter Ladefoged, pages 1–14. Academic Press, Orlando, 1985.

Daniel Jones and Solomon Tshekisho Plaatje. A Sechuana Reader, in international Phonetic Orthography (with English Translations). The University of London Press, London, 1916

Answer 3

The Cardinal Vowel chart is a rough, impressionistic tool. Nobody is claiming that it is a precise measurement of the quality of a vowel but rather that it is "good enough" for the purposes of recording and conveying an impression of the approximate quality of a vowel.

Although it is true that as far as normal, spontaneous speech processing is concerned, listeners may be desensitised to subtle variations in vowel quality around the targets of phonemes in their native language(s), this doesn't necessarily preclude that with training, phoneticians can still judge the quality of vowels precisely enough for the purposes of the Cardinal Vowel chart.

The CV chart isn't very well suited to studies relying on precise measurements of vowel quality. For such studies, it is more common nowadays to take an approach similar to that outlined by JoFrhwld, in which vowels (or indeed, continuants in general) are analysed for their formant frequencies. With some judicious mathematical jiggery-pokery, you can just about contort formant frequency measurements into something resembling a CV chart. But there's not always much motivation for doing so: once you're analysing in terms of measured formant frequencies rather than impressionistically, you may as well just plot those measurements on some graph or other, even if it doesn't quite look like a CV chart.

So the bottom line is: if you are asking "how do I find the precise position of a vowel on the CV chart?", that's probably an indication that the CV chart isn't suitable for your purposes.

Answer 4

There has long been a secret war in the phonetics community: lets call the two camps "articulators" and "acoustics". As the names imply, there are two competing ways of describing vowels: what the mouth does (position of tongue, jaw, shape and position of lips etc., this is reflected in the IPA) and what the vowel sounds like, which is visible in a spectrogram for instance. With the latter one uses the formants to map the sound into a vowel diagram, and funnily enough if you connect the dots it still looks like a skewed square! These days, measuring how the mouth articulates a sound is hopefully done with for instance an MRI-machine and not the old way, which includes mirrors, videocameras, painting the mouth, very strong x-rays, casts of the mouth etc. Being a phonetics informant in the old days could be dangerous to your health. Analysing the sound is dirt cheap in comparison.

Without knowing which method was used you can't really compare vowel diagrams.

Answer 5

Regarding the vowel charts in the question, it's important to be aware that these show the phonemic vowels of the languages. When placing vowel phonemes on the quadrilateral they are typically spaced out evenly, as we see in those charts. This is especially so when the vowel systems are neatly balanced, like the ones shown. This should not be taken as a claim that these are the precise values of the vowels. All it really claims is that these are the contrasting vowels of those languages and shows how they contrast. For example: from the chart for French I can see that it has an unrounded high front vowel (represented there by /y/) and another unrounded front vowel that contrasts in being slightly lower (represented by /ø/). A trained linguist, even if they had never before heard French (another weird and exotic IE language I believe), could use these as a rough guide and probably make a fair stab at pronouncing words containing these vowels.

As described in other answers to this question, locating vowels precisely on the quadrilateral requires acoustic analysis. This page contains examples of vowel quadrilaterals based on auditory judgement and ones using measurement of formant frequencies.

Answer 6

As someone who does language documentation and isn't too crash hot at analysing formants there is another way to analyse vowels - which is based on the perception of the listener rather than the production of the speaker.

The easiest way to do this is to find minimal pairs or sets - so in English we know that /a/ and /i/ are separate vowels because people have the words "pat" and "pit" so they're separate vowels. You can also do this for consonants - "pat" and "bat" are different words so /p/ and /b/ must be different.

Of course, this measures something slightly different to production - we call the production stuff "phonetics" which sets out to give an inventory of all sounds but what I'm doing is something called "phonology" which only looks at the sounds speakers can tell the difference between. A quick example - put your hand to your mouth and say "pin" and then "spin." Notice that "pin" gives you a little puff of breath but "spin" doesn't? That's because technically they're slightly different sounds - which someone interested in the phonetics would case about. To speakers though the /p/ in both "spin" and "pin" sound the same so I'd treat them as one sound.