3

Unlike vowel formant frequencies, tones are trickier to determine, since the F0 of a TBU may depend on a lot of factors, such as the speaker's age, gender, mood, etc. Thus, it's quite often the case that accounts based on impressionistic transcription disagree on whether a tone is, say, 24 or 35. I know from looking at Praat that the same tone - even if it sounds perceptually similar to me - can produce completely different-looking F0 contours in Praat.

Is there any established experimental procedure to determine the pitch(es) of a tone, either in Chao's system (1 is lowest, 5 is highest) or the Central Americanist one (reverse)? Thanks!

3
  • I think this question needs to be clarified. What is the input? Just a sound file? Is it tagged or labeled in any way? Is it a whole utterance made up of many words? Is it a single syllable in isolation? Also, it's not just speaker characteristics that change the F0; it's the actual linguistic context--utterance position, focus (emphasis) condition, sentential intonation (e.g., question vs. statement), etc. – musicallinguist Jan 24 '17 at 22:45
  • I was asking for an entire experimental procedure rather than just an analytic technique, so I didn't specify any particular type of input to allow for more freedom in the answers w.r.t. experimental setup. Yes, I'm aware that there are many more factors affecting the F0 - this is what makes tone tricky! D: – WavesWashSands Jan 25 '17 at 3:49
  • But I think perhaps word lists, balanced enough with respect to surrounding environment, segments, etc. would be what I 'd do, rather than single syllables in isolation. Consultants sometimes (often?) don't explicitly know what the register tone actually is... – WavesWashSands Jan 25 '17 at 3:56
3

There is no (implementable) experimental procedure for doing this, and the main reason is that the thing you want to do isn't well-defined (but I'll try to define it sort of, with the goal of making such an experiment be useful). There is a trivial way to get F0, so in that sense you can always come up with a numeric value. But what you want is a reduction of continuous pitch to certain ranges, let's say 5 intervals (it should be 6 for purposes of linguistic contrast, but whatever). One way to do this is to divide the possible range of human voices into 5 bands. However, it is a fact that in some languages, pitch range tends to be high and in other languages it tends to be low, so you could come to the absurd conclusion that all tones are 1 in Chinese and 5 in Bari.

An alternative would be to compute an observed range for a language (90% of all F0 values in a corpus fall within that range), and then divide into 5 bands. You still get the same absurd result that speaker A has only tone 1 and speaker B has only tone 5. You can continue to narrow the subject pool, so that you only look at one speaker and determine his/her range, and then mark individual words for that speaker in terms of their absolute pitch numbers (scaled down to 5). You might (should) also use a non-linear scale to map F0 to numbers, because a 10 Hz difference at around 100 Hz is perceptually much bigger than a 10 Hz difference at around 400 Hz.

This doesn't tell you anything about tones, though – you'd need an independent determination of what the tones of the language are (tone is a phonological, contrastive concept). There are a lot of physical properties that go into tone: not just F0, but also movement of F0 and phonatory properties (especially in SE Asian languages). Still, you can set that aside, if the interest is just pitch. Mandarin is a good example of the ambiguity of tone counting: it has 4 holistic "tones", and 3 levels. This is pretty common, that you have to specify the target points (H, L, M... what the numbers are about), and the transitions (rising, falling, level). The linguistic numbering systems do that by specifying e.g. "52" meaning starts at level 5 and ends at level 2. But you have to first do a phonological analysis, to determine if there is a "52" tone in a language that contrast with a "51" or "42" and so on.

Suppose you have the analysis, and you know there is a H, L, M, and contours ML and LH. Now the goal is to turn those tone categories into something closer to physical, namely the pitch numbers that you got by dividing H0 into 5 bands. The more bands you have, the more accurate your tone-to-pitch mapping will be, but let's stick to 5. You have to parse the stream of pitch integers to tone-marked syllables so that you know for instance that the pitch of a given ML syllable is (using "5 is highest" numbering) "4 3 3 2 2 2 1 1 1 1". This comes from directly scaling F0 measurements as you get them from Praat, for instance (evenly spaced). Now you need to compute the "first" and "last" pitch value, since the goal is to reduce the pitch trace even more, to just two specifications. There is a huge problem with figuring out a meaningful answer to the question "what is the initial pitch". Do you adjust away from consonants to avoid pitch artifacts (lowering after voiced obstruents, raising after voiceless)? Do you just take the first and last points and assume that the bulk of points in between aren't actually essential? Do you divide the points evenly and compute the mean? (If the syllable has 3 tone points, get the mean in thirds; if it is level take the mean of all points)? I'd suggest the latter: divide the pitch points evenly according to how many phonological targets there are. Round the results. Then you can say that in the aforementioned syllable, there is a 31 tone, implementing the category ML.

What this basically does is massively reduces the data which is a raw F0 trace, to something more symbolic. The next question would be, and what does this show you? It would be inferior to an actual pitch trace, in obscuring many important details of tone implementation (remedy: expand the range of tone integers to 10). It still requires you to know what the phonological tones of a language are, although it does solve the "Chinese" puzzle that a falling tone can be judged to be 53, 42, 52 and various other numbers (and: it is documented that different researchers into Chinese actually assign different numbers for the same dialect).

As an experiment related to understanding pitch perception, this could be extremely useful if done systematically and with sensitivity to the variation that exists in human languages. Or it could be used as a Chinese-specific calibration technique. AFAIK, the absolute-number approach used in Chinese studies does not exist in Meso-America, so if a language has 3 levels then only 1,2,3 are used.

When you get to languages with upstep and downstep, the aforementioned system completely collapses since there is no upper bound on the number of "categories". If you do decide to deal with upstep and downstep, you can take advantage of the fact that the range itself changes over time. Initially, the range might be between 200 and 150 Hz, but at a point (where there is a downstep), the range lowers to be between 175 and 125 (not the vast oversimplification here). If the size of the range is, say 50 Hz, and you're computing 5 pitch intervals, then the F0 to pitch-integer conversion is (f0-b)/5, where b is the baseline value at the low end of the range, and downstep means "reduce the baseline value". The problem is that you need to determine that reduction empirically (and it isn't likely to be a constant).

4
  • I see. I guess I asked the question with the assumption that a good phonological analysis should be conducted after a good phonetic analysis, but I guess this isn't the case when dealing with tone. I mainly asked the question because I thought it would help to know the pitches used in register tones in great detail so that the relationship between register and sandhi tones can be more easily explored, and I think applying your procedures will help with that! – WavesWashSands Jan 25 '17 at 3:52
  • 1
    A phonological analysis has to go hand-in-hand with a phonetic analysis: each progresses incrementally. You have to know a little about impressionistic contrasts to be able to learn a little about phonetic detail, so that you can learn a bit more on the next pass. – user6726 Jan 25 '17 at 5:13
  • 1
    Maybe I should have said something specifically focused on Chinese, in case that is relevant to you. I admit that Chinese tone baffles me, but I have come to believe that clarity will only come from looking at sandhi relations. It is standard to talk about e.g. Xiamen 24 → 22, 22 → 21, 44 → 21, but I think you have to start with something less a prioristic – A → B, B → C, D → B. That is, Chinese really challenges phonetically-determined theories of tone. – user6726 Jan 27 '17 at 0:04
  • Yes, that's what's relevant to me as you've guessed :) I agree with your judgement that starting with something less aprioristic is a superior approach, one that I would have benefitted from using. The tone sandhi phenomena make a lot of sense when you read their analyses from the literature (incl the Xiamen one! such a neat cycle), but when you're working on a scarcely documented dialect, it gets so much more baffling, even if it's a Southwestern Mandarin dialect with mild sandhi compared to Wu and Min. :P – WavesWashSands Jan 27 '17 at 15:45

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.