Linguistics 431/531 - Phonetics
Plotting your own vowels
Homework Assignment 6 - due 11/5/15, in class
Name:
Date:
Vowels are surprisingly variable across languages, dialects, and speakers. While vowel quality is
one of the most important characteristics that we use to distinguish between dialects of English,
individual speakers can vary greatly as well. The idiosyncracies of how you speak are a product
of not only where you grew up and the language you heard, but also of how your peers and
parents spoke to you.
Since we know that vowel quality can be approximated by the acoustic signal, the purpose of this
assignment is to plot your own vowels in the acoustic space. At the end of this assignment, you
will have learned something not only about how you produce your vowels, but about how to
record and analyze acoustic data using Praat.
The necessary steps of this assignment are for you to record yourself saying some words
illustrating different vowels, produce a graph of your vowel space, and answer some questions
about the acoustics of vowels.
Let's get started:
1.
Record the English words below into Praat. Open Praat, select "New: Record mono sound"
and a dialog box will appear. Press "Record", record yourself producing two repetitions of
these words, and then press "Stop." Now click "Save to list" and close the dialog box. You will
see a new sound file in the Praat Object window.
If English is not your native language, it would be preferable for you to substitute similar
example words from your own native language. If you do this, try to find words that begin with
no consonant, a glottal consonant (/ʔ, h/), or a bilabial consonant (/b, p/), and that either have no
final consonant or the same final consonant, e.g. [hab, hib, hub] or [pa, py, pi]. If you find that
you have too few vowels in your language (5 or less), perhaps try producing vowels in one
context (after bilabials) and in another as well (after alveolars) and compare them.
Here are the words to record in English:
heed hid
[i] [ɪ]
aid head had odd awed
[eɪ] [ɛ] [æ] [ɑ] [ɔ]
hud herd owed
[ʌ] [ɚ] [oʊ]
hood
[ʊ]
who’d
[u]
2.
Once you record these words into Praat, you can investigate what their formant values are.
Select your sound file and click "View and Edit." You may have to zoom in to see a
spectrogram. If you do not, make sure "Spectrum: Show Spectrogram" is selected.
1 3.
Select the first repetition of each word and zoom in to the word. Place your cursor in the
middle of the vowel portion. You should be able to distinguish the aperiodic noise in [h] from the
periodic vowel which follows it (though, to tell the truth, the boundary between these two can be
rather gradual). From this midpoint, you should see dots which correspond to the formant values
for that vowel. If you do not, make sure "Formant: Show Formants" is selected. You can
estimate the formant values at this midpoint by doing either of the following:
a.
Try to center your cursor on the dot corresponding to the formant. You should see a value
on the left of your window with the estimated frequency value. You are, in essence "eyeballing"
where the formant lies.
b.
Select "Formant: Formant listing" from the menu. You will be shown a info box which
contains the exact formant estimates at the midpoint where your cursor lies.
Keep a table of vowel formant measurements for each of the vowels in the word list. Note that
no formant tracker is accurate more than 5-10 Hz, so it doesn't make sense to include numbers
after the decimal point, i.e. just record "562 Hz", not "562.3285."
First repetition:
heed hid
F1
aid
head had
odd
awed hud
heard owed hood who'd
F2
F3
Now, do the same thing for the second repetition of these words:
heed hid
aid
head had
odd
awed hud
heard owed hood who'd
F1
F2
F3
4.
Now that you have your formant values, plot the vowel formant measurements on an F1 vs.
F2 chart like we saw in class. You can do this by using Microsoft Excel or some other graphics
package (e.g. the free statistics package R [www.r-project.org] has a simple plot() command for
this). Try plotting the vowel formant measurements on both linear and logarithmic axes, and be
sure to label the points in the plot with the IPA vowel symbols. You can do this part by hand.
2 Then again, you can plot everything by hand if you use a piece of graphing paper and trace the
approximate locations of formants. Sometimes the old-fashioned methods are easiest.
5. Answer the following questions (on a separate sheet of paper):
Q1: Is there any advantage to the logarithmic scale? If you have an intuitive or
auditory/perceptual notion of the shape of the vowel space, which type of display conforms more
to that intuition?
Q2: Does omitting F3 in these plots cause any important vowel-specific information to be
neglected? If so, specify what might be lost (don't guess - look at the values you got from your
own vowels).
Q3: Which vowels have the most mouth movement over the duration of the vowel? To answer
this you have to go back and stare at the spectrograms.
Q4: Compare your first repetition to the second. How different are your formant values? Which
vowels differ the most? Which formants differ the most? What does this suggest about your
acoustic vowel space? What might you expect if you made more repetitions of each word?
In addition to answering these questions, turn in your table of formant values and your
vowel plots.
3 Praat tutorial (for spectrograms/formants):
1. Spectrogram settings. How to get there: View and Edit → Spectrum → Spectrogram settings.
For vowels you'll want to look at a “view range” from 0 to 5000 Hz. I find that spectrograms
look better with a somewhat longer window length and a smaller dynamic range, than is given by
default. The smaller the window length, the more harmonics are visible in the spectrogram. The
dynamic range refers to a display setting - the spectrogram will only show the highest 50 dB of
the signal intensity here.
2. Formant settings. How to get there: View and Edit → Formant → Formant settings.
4 You should probably change your "maximum formant" to match your spectrogram view range.
However, the main parameter here than can really change how well the automatic formant
tracker works is the number of formants. Look at the spectrogram and count how many dark
bands you see. With a frequency range of 0 to 5000 Hz the number will generally be about 4 or
5. As a rule of thumb, we expect to find about one formant for every 1000 Hz, though this will
depend on the speaker and the vowel. For vowels with formants very close together, telling the
software to look for more formants than this might help to distinguish two very close bands.
At the very least, this should teach you that "default" settings do not work for every vowel or
even every speaker. For instance, taller speakers with longer oral cavities have lower resonances
and thus formants which are closer together than shorter speakers have.
On the left, I have set the number of formants to "4", whereas on the right, I have set the number
of formants to "5." Notice how the first and second formants are not tracked very well in the
latter half of the vowel on the left (as they get closer). This is improved when I increase the
number of predicted formants. So, be careful and don't necessarily trust the formant tracker to get
the right answer – you have to judge – is it correctly putting the formant tracks on the formants?
===========================
Tutorial: Using the R package to make nice looking graphs.
You might not like R because it is a command-line language (but you might be into this too).
This means that you have to type commands to it instead of point and click. But it does make
nice graphs and it can do any statistical analysis known to man! Download R for free from
www.r-project.org and open it up.
Here are the commands that I used to make graphs of some formant measurements. You will
want to include more data and vowels here for plotting your whole vowel space. Just add F1 and
F2 values to each of the commands below along with the IPA characters in the "names" line.
# my data
> F1 = c(300, 500, 800, 500, 300)
> F2 = c(2400, 2300, 1300, 1200, 900)
> names = c("[i]","[e]","[a]","[o]","[u]")
(If you want to nerd out here, you could print any IPA character by entering in the 4 digit
hexadecimal for the character in unicode, e.g. [\u025B] will print as [ɛ]. Though, I'll leave it to
you to look up unicode hexadecimals for IPA. The program IPA Palette on the mac is useful for
this - http://www.blugs.com/IPA/.)
5 # my linear plot
> plot(F2,F1,xlim=c(2500,800),ylim=c(1200,300), type="n")
> text(F2,F1,names)
This produces the following:
[u]
400
[i]
[o]
800
[a]
1200
1000
F1
600
[e]
2500
2000
1500
1000
F2
6 Now, to plot things logarithmically, I used the following commands:
# my logarithmic plot.
> plot(F2,F1,xlim=c(2500,800),ylim=c(1200,300),type="n",log="xy")
> text(F2,F1,names)
[u]
400
[i]
600
[o]
800
F1
[e]
1200
[a]
2500
2000
1500
1000
F2
These are the plots for one particular speaker, but yours may look quite different. If your values
fall outside of the range in the axes in the figures, adjust the range in the commands and rerun the plot() commands. You'll notice that the ranges here are called "xlim" and "ylim" and
the values inside each specify the range of the axis.
7