Identifying phonemic boundaries in Praat

Question

I am trying to segment some connected speech in Praat, and want to get the boundaries between phonemes as accurate as possible. I am finding that in many cases, one sound blends into another and it's hard to say where the dividing line is.

Praat obviously provides a good few clues, but I'm not sure how much weight I should attach to them. Looking at the word /cə/ for example:

The waveform has a very clear change in shape near the beginning, so the obvious starting point is to treat the first shape as /c/ and the rest as /ə/ - but if I do that the bit that is supposed to be /ə/ sounds like /də/.

At a slightly later point, I would say that the formants settle (they are a bit wobbly throughout) - but again a boundary there gives me /də/.

Slightly later still, the first vocal pulse appears. I am not sure how far I can trust the vocal pulses shown in Praat, but in any case I still get the /də/ if I put the boundary here.

If I move the boundary so that the second segment sounds like /ə/, the end of the bit that is supposed to be /c/ is audibly voiced.

Clearly there are a number of things going on in this transition and they don't all happen simultaneously (or instantaneously).

Is there a principled basis for choosing one of these points - or perhaps another one altogether - as the boundary between consonant and vowel?

Answer 1

There are enough criteria that the decision is unprincipled, that is, there isn't some unquestionable principle that you can use to deduce where the lines must go, if you are looking for phoneme boundaries. For example, in parsing English [kʰɑɹ], you have to decide whether the velar release burst is the end of the consonant, or do you go for the beginning of voicing (which itself calls for a judgment: do you need one complete semi-sinusoidal period in the waveform to determine that you now have voicing? do you subject the parsing to a stronger criterion of full modal voicing?). This was a topic that was central to phonetic research in the early period (late 50's to early 70's), so reading linguistic phonetic works from those days will be very informative. Two examples are Peterson & Lehiste 'Duration of syllable nuclei in English' and Lehiste 'An acoustic-phonetic study of open juncture'. This class handout gives an overview of main considerations, and would be a good quick first read.

The best approach is to provide more information, not less, and then "discover" the most stable measures after the fact. However, when it comes to V-C and C-V formant transitions, it is better to let your eyes provide the answer rather than your ears. The consonant closure properties (formants) will be detectable for quite some time after the release of the consonant constriction, because the tongue (e.g.) cannot instantly snap into the target vowel configuration. Yet there will clearly be properties that identify that period as being a vowel, not a stop. This becomes a problem when parsing vocalic sequences, that there's a huge uncertainty in parsing [awa] and the like. I don't know if there is a "majority" approach among phoneticians: I use a "split the difference" approach where the glide boundaries are the midpoint in the stretch that isn't clearly just [u] or just [a].

Answer 2

If you base your segmentations on your perceptions, you will never make a good segmentation.

It is normal to hear [de] when you play only [e] because [d] is a stop, so a "silent" sound. As [e] is preceded by a silence, so you get a sound it looks like [de]: [d] + [e] = silent + [e].

You should only trust the graphic representations of sound (signal, spectrogram, ...). Even with these graphics, it is not easy to find the phonetic units, because of the coarticulation. If you are not interested in the duration, the boundaries of your segment are not so important (few ms' of differences can be judged as negligible), while you embrace the steady part of your phonetic unit (when the articulators reach the place of articulation).

Answer 3

Personally, I treat transitional phenomena, like burst or palatalization, as part of the consonant. Vowels normally have more or less stable formants, so if you want to get the formants, you need a clean period to analyze.