Overview

Part 1:

• Assumptions
• Why (instrumental) phonetics?
• Acoustic data
• Using Praat

Part 2:

• Types of measurements
• Making figures
• Numerical and tabular data

Part 3:

• Beyond Praat: articulatory data
• Practical considerations
• Open discussion period

Regular breaks will be interspersed

About the slides

These slides are a web page

• Use right, left arrow keys to navigate (or click to advance)
• Press “A” to see all slides at once, and “A” again to go back to slide view
• Links are formatted like this
• References look like This (1985)
• All references have links provided in the bibliography

The slides are hosted here on GitHub

• Along with all associated media
• This slideshow’s URL is mfaytak.github.io/afriphon/btd-1.html

References

Nearly all references mentioned during the workshop are linked at the end of the slides

• I have prioritized open-access materials as much as possible
• I have aimed to include many “model papers” which demonstrate use of new concepts
• Attendees are encouraged to look up these sources at the end and reinforce what they’ll learn today

The situation

Africa has long been regarded as central to phonetic description Ladefoged (1968), Maddieson & Sands (2019)

• Complex tone systems abound
• Clicks and tongue root harmony are found virtually nowhere else in the world
• Most African languages remain phonetically undocumented Whalen, DiCanio, & Dockum (2019)

But relatively little of this scholarship involves African scholars

• An issue we’ll revisit a few times

Participants

We presume that you, the participants, are:

• Employed by, or trained at, a university on the African continent (sub-Sahara)
• Researching in a low-resource context
  • External grants are uncommon
  • Institutional support is low
• Familiar with phonetics in theory, but not necessarily in practice

If you are not in this group, we ask that you prioritize those in this group for questions and feedback

• Unlike a typical talk, we encourage people to ask questions at the end of any slide

All of us

Let’s assume that we are all committed to:

• Improving empirical coverage of African languages’ sound structures
• Building speech data resources for African languages
  • Ethos: “some data is better than no data”
  • Crucial to advancing scientific research
  • Important as starting point for technical development
• Developing capacity for an African phonetics practice
• Doing these in a way which is adapted to local needs and demands

Definitions

This workshop is an introduction to instrumental phonetics

• Relying on instrumental readouts for analysis
• Not exclusively impressionistic: using the ear and transcribing speech sounds

“Phonetic” work may also refer to non-contrastive subphonemic detail

• The specific way a phonemic contrast is produced
• This is also important here, but we are focusing on instrumental methods

Why instrumental phonetics?

Several practical advantages over impressionistic approaches

• Neutrality in the face of analytical and perceptual bias
• Precision and reliability in detecting contrasts
• Community use of the created data

Neutrality

In impressionistic phonological description, all presentation of data is filtered through the worker’s theoretical analysis

• For example: autosegmental representations often make transcriptions more abstract Hyman (2014):545

Phonetic recordings allow better testing of hypotheses about phonological structure

• Recordings do not intrinsically involve an analysis, and can be reanalyzed at a later date
• Transcribed data (being analyzed) is much harder to use for this purpose

Neutrality

Even a trained phonetic ear is prone to making occasional mistakes based on perceptual bias

• For example: nasal consonant codas are more often misidentified after non-low vowels Zee (1981)
• Transcription mistakes permanently enter the record

Precision

Not all contrasts can be easily described by the analyst’s ear, and especially transcribed quickly in the moment

• Fine vowel contrasts (especially central vowels)
• Diphthongs versus consonant secondary articulation
• Prenasalization versus N+C clusters
• Subtle differences in tone level and contour
• Multiple downsteps/upsteps

Recordings allow for careful listening later

Community use

Recordings are required for instrumental phonetic work: many incidental benefits

• Speaker community may access the data
  • Literacy development (teaching tools)
  • Technical development (speech resources)
• Community of scientific researchers may access the data
  • New analyses
  • Comparative work
• Analysis may be replicated later by other researchers (or the same researchers)

Complementary methods

The aim is not to displace impression-based methods, but to complement them

• Transcription will always be needed at some level
• Our point is that it should not be exclusively relied on as the analytical object
• Whenever possible, transcriptions ought to be supplemented with recordings, visualizations of recordings, or instrumental measures as evidence
• Instrumental measures as “second opinion” for analysis

Desired qualities

We always want acoustic speech data to be:

• Low in background noise
• Sufficiently loud against background noise, but not too loud
• Free of echo

Certain details of format are also important:

• Record using a high sampling rate, at least 22.1 kHz
• Save in non-compressed format (such as .WAV; avoid .MP3)

Good recording

Here is an example of a good recording

• “La plume de ma tante”
• Speaker’s voice is much louder than background, but is not too loud
• Background is free of avoidable noise
• Practically no echo

The following slides contain recordings which fail on one of the points above

Too noisy

Recordings should not contain excessive background noise

• Continuous noise from a fan
• Intermittent noise from touching the microphone

Any noise, however quiet to your ears in the moment, will be much louder in the recording later

How to improve

Listen carefully to your surroundings, and avoid:

• Rain on the roof (especially metal roofs)
• Appliances (refrigerators, any motors or fans)
• Busy roads (trucks, taxis)
• Chickens, goats, children, etc.

Speaker should also minimize non-speech noise:

• Touching or scratching microphone, or contacting shirt collar
• Producing background noises when emphatic (striking chest or table, clapping hands)
• Phone ringing or vibrating

Too much echo

If echo is strong, speech ends up overlapping itself; problem for listening and analysis later

• Slight echo (in tiled hallway)
• More echo (in concrete stairwell)

How to improve: listen for echo and choose surroundings which have less

• “Soft” rooms reduce echo (couches, carpets, pillows, hanging clothes); tile, stone, and cement produce echo
• Record in the back seat of a car (motor off) if available
• Record outside if no suitable room exists

Too loud (clipped)

If the speaker is too loud and/or too close to the microphone, the device cannot respond enough; clipping results

• Clipping of whole utterance
• This can also happen for stops and fricatives only, where the releases “pop” in the microphone

How to improve: make test recordings after you position your microphones

• If there is general clipping, microphone needs to be further away or speaker needs to be quieter
• If stops “pop”, position microphone to the side of the mouth
• Gain can often be adjusted if you are using a recorder

Equipment

At the most basic, you only need something to make recordings on. One of the below:

• Laptop computer with Praat
• Smartphone with recording app
  • “Awesome Voice Recorder X” (free with ads) is a good app
  • Others must definitely exist
• Hand recorder/memo recorder
• Professional recorder (Zoom H4N, etc)

A way of transferring files off of the device and backing up for future analysis:

• SD card
• USB drive or external drive

Headphones, to check recording quality

Equipment

External microphones can increase the quality of acoustic data by recording less echo and background noise

• Look for cardioid or unidirectional in the description if you are recording single speakers
• Recording public events may require an omnidirectional microphone

An acoustic baffle can reduce echo

• Soft, fluffy objects: blankets, sofas, etc.
• The seats of a car
• This can also be achieved by selecting the recording setting carefully, and using a cardioid microphone

Other tips

Do not use computer noise reduction/filtering in general

• Go into your computer’s sound settings and turn this off; make sure apps (i.e. Zoom) don’t have separate filters
• Normal to hear slight “fuzz” in background, or very slight echo

Recording over Zoom works surprisingly well, if all else fails Ge, Mok, & Xiong (2021); Sanker et al (2021)

• Some small effects on the recording are noted
• Turn noise cancellation off if using to record (switch to “original sound”)
• Better still to use the speaker’s phone on their end to record, and use Zoom to supervise the process

Downloading and configuring

Download Praat from fon.hum.uva.nl or praat.org

• Mac: the .dmg file contains Praat.app (which you can run immediately) and the option to install
• If you have issues opening, see 2. “How to Start” here
• Windows: the .zip file contains Praat.exe, which you simply double-click to run (instructions here if needed)

Windows and manual

Both the object window (shown below) and picture window appear when you open Praat (we’ll ignore the picture window for now)

Basic issues can often be solved using the manual

• The manual is also available online here
• Other introductory guides can be found here

First recording

Let’s record ourselves saying a “good day, good afternoon” greeting in whichever language you would like

• Make sure you “save to list” before closing the recorder!

Stored sound

You should now have a Sound in your objects window

• Select the Sound object and a menu will appear to the right
• Click “Play” to hear your sound

Saving your recording

We’ll end this by saving our work as a .WAV file (the standard format for phonetics work)

• The file is not yet saved, only stored in Praat
• If you close Praat without saving the file, it will vanish

Importing a sound file

We might also import sounds which have already been recorded

• Download this ZIP file (673 KB), which contains all of the files we’ll use
• Use the “Open” dialogue to open bbk-hills-of-bushes.wav
• That file contains one utterance of Babanki [tə̀tāŋ tə́ tə́ꜜtóʔ] “hills of bushes” from Faytak & Akumbu (2021)

Viewing a sound file

We’ll start by viewing Babanki [tə̀tāŋ tə́ tə́ꜜtóʔ] “hills of bushes”

• Select the Sound and click “View and Edit” in the right menu

Viewer window has button and keyboard controls

• Select by clicking, SHIFT-clicking, or clicking and dragging in sound
• Press TAB to play selected sound, or click bar below window
• Control buttons (zoom in, zoom out, scroll) are at lower left
• Keyboard shortcuts: CTRL+ (PC) or Command+ (Mac)…
  • A: zoom out to all
  • N: zoom to selection
  • I: zoom in
  • O: zoom out

We may wish to know where the words and segments are, but we lack useful landmarks at this point

• Each “blob” is a syllable
• No indication of tones (yet)

Making a TextGrid

TextGrids are one of the most useful features of Praat: annotate and organize your audio files

• Using the menu as shown below, we’ll make a TextGrid for our Babanki sound object

Interval tiers

Let’s use the interval tier “sentence” and use it to transcribe the utterance

• Contains a list of ranges in time separated by boundaries
• Useful for marking off words, utterances, and some vowels and consonants
• Here, click before the beginning and after the end of the utterance, then type a transcription or translation in the middle

Point tiers

Let’s use the point tier “stop” to mark off where each [t] release happens

• Contains a list of points in time
• Useful for instantaneous events
• Note that intervals are generally more useful for most segments
• Can’t click and highlight range like interval tier

Adding and removing tiers

We might be dissatisfied with how the tone marks are displaying; we could make a new tier for tones (autosegmental style)

Amending the TextGrid using the new tier:

Saving TextGrids

Using CTRL+S or the menu shown below, you must save your TextGrid when you are done

• If you close Praat without saving your TextGrid, it will vanish
• Much like recorded audio files which are stored in the object list

Data displays

We may wish to provide further details in our TextGrids, but we encounter another problem here: how to interpret the data?

• Praat shows waveforms and spectrograms
• Note the simpler TextGrid (segments, tones); bbk-hills-of-bushes.TextGrid in downloaded files

Waveforms

Waveforms show sound pressure (the pressure that sound waves make on the microphone) versus time

• We expect any voiced speech signal to oscillate because the vocal folds open and close in a repeating pattern
• Other voiceless sounds show no movement away from “zero line” and no clear oscillation

Sonority

Sounds produced with a more open mouth are louder and more sonorous Parker (2008) ; these are thicker on the waveform

Spectrograms

We can also show the spectrogram for our recording

• Breaks down our waveform to give us information about the sound spectrum and where its energy is (high or low frequency)

Spectrograms are three-dimensional, and show time vs. frequency vs. sound pressure (color)

• The darker the spectrogram (and the thicker the waveform), the more sound pressure there is

• Think of it as an elevation map (Mt. Cameroon pictured)

Interpreting spectrograms

Vowels, semivowels, and approximants have characteristic striping, horizontally and vertically

• Voicing is visible as vertical stripes in vowels, and as a small “bar” at bottom
• Formants are clear horizontal bands above the voicing bar

Interpreting spectrograms

Nasals look somewhat like vowels, with “smudged” formants and less energy (darkness)

• Due to the opening of the velum: nasal cavity “muffles” the sound
• Much like soft objects in a room muffle echo

Interpreting spectrograms

Fricatives have high-frequency or low-frequency noise

• Dark irregular smudges across entire areas of spectrogram
• The further back the fricative, the lower the average frequency
• Compare [s] and [ʒ]: [s] is further front, has higher frequency, sits higher on spectrogram

Interpreting spectrograms

Stops show an absence of (most) energy followed by a burst across the whole spectrum

Affricates look similar, but as if they were followed by a fricative

• From Babanki [kə̀dʒóm] ‘Babanki’ Faytak & Akumbu (2021)

Prenasalized segments

Prenasalized stops and affricates often have a long nasal closure followed by a short oral closure which we can see clearly in the spectrogram

• Note that most of the spectrogram energy goes away when oral closure begins in [ⁿpfʲ] below
• From Babanki [kə̀ⁿpfʲɨ́ŋ] ‘owl’ Faytak & Akumbu (2021)

Why make figures?

Figures are an easy way to present a small amount of phonetic data in scientific papers

• Give support to phonological judgments
• Support particular transcriptions
• Provide more detail for any especially unusual sounds or sound contrasts

This section: professional-looking and informative data displays

Picture window

Waveforms and spectrograms can be “drawn” or “painted” (respectively) in the picture window

• This also opened when you open Praat; we’ll stop ignoring it now
• The blue rectangle indicates the plot area where the figure will be drawn: change size by clicking and dragging as needed

Drawing a waveform

“Draw” (in the object window) is for any line-based drawings, including waveforms

• Select a Sound to draw its waveform

The result: a waveform drawn within the plot area

Drawing a TextGrid

TextGrids can be drawn as well, using the same menu as Sounds

• By default, they appear with an empty space above them

Combining TextGrids and waveforms

A Sound and TextGrid can be drawn together very easily: simply select both and choose the Draw menu as before

• Adds TextGrid annotation to the waveform drawing

Extracting a spectrogram

“Paint” is for drawing spectrograms and other objects, but: we need the right object to do this

• Select Sound and click “View and Edit”
• Select from the Spectrum menu “Extract visible spectrogram”

Painting a spectrogram

This sends a Spectrogram object to the object list; when we select this we get a “Paint” option under Draw

The result: as expected, but a bit too tall for its width (we might make the plot area wider/shorter)

Spectrograms and TextGrids

Drawing a spectrogram and a TextGrid at the same time is a bit more complicated

1. Paint the spectrogram, but uncheck “Garnish”
2. Add the Y axis marks using the “Margins” menu

3. Add a Y axis label, usually “Frequency (Hz)”

• I’ve also added a box around the plot here with “Margins” > “Draw Inner Box”

Spectrograms and TextGrids

3. Resize plot area to be taller than spectrogram (pictured), and Draw the TextGrid

The result: TextGrid annotations on top of the spectrogram

Saving

As with everything else in Praat, you must save before closing the picture window or you will lose your work

• If you don’t keep the plot area in the same place (covering your figure or margins), you won’t save the entire figure

Why numerical measurements?

Now we’ll turn to taking numerical measurements in Praat

• Actual calculation
• Displaying figures of these measurements
• Storing measurements as tabular data (at the end)

More than showing an entire sound file as a waveform or spectrogram, focusing on a specific phonetic property can be useful

• Make a figure for this specific property
• Focus on topic of interest for your discussion
• Display with a TextGrid, a waveform, etc.

Why numerical measurements?

Also lets us measure many utterances and summarize, which also allows us to handle phonetic variation

• Languages differ in their phonetic implementation of the “same” segments
• Speakers of the same language produce it differently depending on their history, social stance, etc
• Even phonologically identical words can differ slightly phonetically Gahl (2008)

Because of this it’s best to collect many observations and average or model the data to remove noise and variation

• Multiple speakers
• Multiple repetitions
• Multiple words

Duration

One of the simplest measures: duration of segments or words

• Simply ending time of the interval minus its starting time (t₂ − t₁), in seconds
• Two ways to acquire this in Praat:
  • Displayed in viewing window when you select an interval
  • If you use the “Query” menu, a text box appears which you can copy-paste the value from

Model use

An example from Babanki (bbk-prenas.wav): proportion of prenasalized consonants which is nasal Faytak & Akumbu (2021)

• Not uncommon for continuants to stop up when prenasalized, e.g. /ⁿz/ > [ndz]
• But /ⁿz/ and /ⁿdz/ (etc.) are contrastive in Babanki
• Babanki speakers produce shorter, lighter prenasalization before continuants
• This may help to avoid merger of continuants and non-continuants

Model use

Other phenomena studied with duration (not exhaustive):

• Voice onset time McKinney (1990); Connell (1994, 2002); Monaka (2006)
• Compensatory lengthening Hamann, Miatto, & Downing (2019); Danis (2020)
• Any other timing relation you can think of (many, many possibilities)
  • Consonant gemination
  • Long vowels
  • Lengthening of segments or syllables due to stress assignment

Pitch (fundamental frequency, f0)

Pitch (and all other measurements which we’ll discuss) have a dedicated menu

• Turn on “show pitch” to use all other options
• f0 track appears over spectrogram (below) when “showing”

Resulting pitch track:

Pitch settings

Settings for pitch are important

• If pitch range is not suited to the speaker, fake jumps in pitch can appear
• Below, the low-falling tone has a fake jump, well above H for this speaker!
• From our prior knowledge of the language we should overrule Praat

Pitch settings

A quick change to pitch range fixes the jump issue

• In “Settings…” under the Pitch menu
• The pitch range was too high for the speaker, so we lower the pitch floor

• This handles the low tones better

• If jumps don’t go away, the “Advanced pitch settings…” window offers some options

Model use

All tonal phenomena involve f0

• Lexical tone, downstep, downdrift, pitch reset, tone depression, lexical tone contrasts
• Intonation (question formation, etc)

Babanki example in bbk-schwa.wav:

• Prefixal L ( və̀-lɨ́m “males”) behave differently from stem L (bə̀lâŋ “plank”) when between two Hs; third word is bə́lə̀ŋ “groundnut”
• HL(prefix)H raises to mid: HM(prefix)H Akumbu (2019)

Model use

Sub-phonemic effects of intonation on tonemes can be examined figure from Rialland & Aborobongui (2016)

• Boundary tones (%) indicate sentence type
• Effect of L% on lexical H in the example below

Other tone topics (not exhaustive):

• Tone depressor consonants Traill, Khumalo, & Fridjhon (1987); Chen & Downing (2011); Mathes & Chebanne (2018); Lotven & Berkson (2019)
• Formal representation of tone Akumbu (2019); Gjersøe, Nformi, & Paschen (2019); Myers, Namyalo, & Kiriggwajjo (2019); McPherson (2020)
• Downdrift, downstep, and pitch reset Genzel & Kügler (2011); Hamlaoui & Makasso (2019); Oppong (2021)

Intensity

A measure of loudness, measured in decibels (dB)

• Intensity menu and settings are exactly parallel to Pitch menu and settings

• The yellow intensity track can be hard to spot
• Higher parts of track match darker parts of spectrogram and thicker parts of waveform

Model use

Intensity is useful for measuring degree of constriction

• Degree of consonant constriction (or lenition)
• Prosodic factors such as stress (in many languages)

Babanki example: stressed (stems) and unstressed (prefixes) don’t seem to be differentiated by intensity

• But stem H has a longer duration

Model use

Implosive vs. non-implosive voiced stops have different intensity profiles figure from Nagano-Madsen & Thornell (2012)

• See also Naidoo (2012)

• Numerous other uses in prosody

f0 and intensity figures

f0 (pitch) and intensity tracks can easily be Drawn to figures as seen above

• In “Pitch” and “Intensity” menus, there is the option to Draw directly to the Picture window with or without a TextGrid
• Best to combine with a TextGrid whenever possible
• Below: pitch contour for Kejom [tə̀tāŋ tə́ tə́ꜜtóʔ] “hills of bushes” from Faytak & Akumbu (2021)

Formant frequencies

Formant frequencies provide vowel quality and other contrasts

• Formant transitions show place of consonants
• Lateral and nasal quality, etc.

Formant frequencies

Turn on “show formants”, and formant tracks for the first three formants (F1, F2, F3) appear

• All have some inverse relationship with a property x: the higher the frequency, the less x
• F1 inversely relates to height (higher F1 = lower vowel)
• F2 inversely relates to backness (higher F2 = fronter vowel)
• F3 inversely relates to retroflexion (low F3 = more retroflexion) and other qualities

Formant settings

Estimating formant frequencies requires calibration for every individual speaker: low pitched voices need different settings compared to high pitched voices

• Often simplified to “men” versus “women”, but there is a lot of variation in pitch for each
• Children (with very high f0) can be especially difficult

The default settings work well for higher-pitched voices

• Lower-pitched voices need fewer formants and a lower maximum frequency

Formant tracking

The default settings of 5 formants in 5500 Hz (for higher-pitched voices) don’t work well for the very low-pitched voice of the Babanki speaker:

Better: changed to 4.5 formants, 4200 Hz; much lower frequency range

Formant figures

Formant tracks work best in a plot of F1 against F2 (which is quite hard to make in Praat), but formant tracks can be drawn like any other measure

• Draw separately from TextGrid and “un-garnish”
• Much like we did for the spectrogram + TextGrid combination
• “Speckle” under “Draw” produces this from an extracted Formant object

F1-F2 scatterplots

F1-F2 scatterplots have F2 on the x axis, F1 on the y axis, with both axes reversed

• Praat does have a limited ability to make scatterplots
• Must manually specify plot size
• Can only use these four point types: x + o .
• Limited ability to label or make legends

Model use

Better scatterplot figures can be made using tabular data (we’ll discuss later) figure from Faytak & Akumbu (2021)

• Here, ellipses indicate where each cloud of points is centered
• With this plot it becomes clear why we reverse the axes: vowel space resembles the IPA vowel trapezoid

Model use

Formant data is useful in figuring out harmony systems; ATR or otherwise

• Dagbani right-to-left ATR harmony also affects the low vowel /a/, counter to previous descriptions figure from Hudu (2016)

• Moro’s height harmony system may condition ‘high’ and ‘low’ /ə/, previously thought to be transparent to harmony Ritchart & Rose (2015)
• Note presentation of data in a table in addition to a scatterplot

Other uses for formant measures (non-exhaustive):

• General description of vowel inventory Koffi (2018), McPherson (2020)
• Characterizing harmony systems Starwalt (2008); McCollum & Essegbey (2020)

Voicing (“Pulses”)

Praat also detects measures relating to voicing: these are grouped under the unintuitive name “Pulses”

• So called because voicing produces repeating, pulsing sounds
• Detecting voicing = detecting regular pulses in the sound signal

The Pulses menu contains the same “showing”, measure-getting, and drawing functions as other menus

The result: each detected “voice pulse”, shown over waveform (not spectrogram)

Display pulses

The pulses themselves can be plotted with a TextGrid like any other similar object

• Shaded areas indicate area where voice pulses occur repeatedly
• Functions like a “voicing detector”

Voicing report

If voicing has a predictable timing but varies in extent, the voicing report may be useful (access in the Pulses menu)

• Specifically the number after “Fraction of locally unvoiced frames”

Model use

Voicing in unexpected places is common for labial-velars; this can be confirmed by looking for pulses figure from Connell (1994)

Other voicing-related topics:

• Mixed-voicing consonant clusters in Taa (“Khoi-san”) Nakagawa (2008); Naumann (2016)
• Lenition or devoicing of consonants Solé, Hyman, & Monaka (2010); Boyer & Zsiga (2013); Bendjaballah & Le Gac (2021)

Getting out of Praat

Praat is useful, but it has important limitations

• Hard to produce all types of figures
• Can’t do statistical analysis (which is standard for phonetics)

Because of this, we often need to export data from Praat into other programs

Tabular data

The most effective way to export numerical measurements: tabular data, that is, spreadsheets

• One observation (time point, segment, etc) per row
• One measure per column
• Name columns using the first row
• Other columns give non-numerical information (speaker ID, segment, word, etc.)

Google Drive or Excel both work well (.xls, .txt, or .csv format)

• I will be using Google Drive for examples, but all of this can be done with Excel too

Data construction: duration

We’ll use duration to demonstrate the most basic spreadsheet construction

1. Select an interval
2. Generate text output in Praat (“Query” menu)
3. Copy-paste output into spreadsheet
4. Add metadata

Repeat this process for each of your observations, starting a new row each time

Praat text output

Text output is generated using the menus above the data display

• Duration: under “Query” after selecting an interval
• Pitch, intensity, formants, voicing (“Pulses”) are all under their expected menus

For any measure other than duration, menus behave slightly differently depending on whether your cursor has selected an interval or not

• If you have selected an interval, you will get a list or an average (mean) of measurements
• If you have not selected an interval, you will get a point measurement

A shortcut exists to get the midpoint of any interval

“Get…” versus “Listing…”

“Get…” provides a single measurement, or a single mean measurement across an interval

• “Get pitch”
• “Get intensity”
• “Get nth formant” (F1, F2, etc)

“Listing…” provides every single measure within an interval

• “Pitch listing”
• “Intensity listing”
• “Formant listing”
• (“Pulse listing”: timing of pulses)

Exceptions:

• Minimum and Maximum Pitch can also be calculated for an interval
• “Voice report” for an interval generates a large number of different measures

Data construction: single pitch

Mean, max, and min f0 measures (from “Get…”) can be copy-pasted into their own cell just like our duration example

Listings (f0 over time) are trickier to paste:

• Multiple columns (time points need to be included)
• Number of time points varies
• May have unsuccessful measurements (“–undefined–”)

Data construction: pitch listing

We’ll take pitch as an example, but these instructions will work for any listing

1. Select interval
2. Generate text output (“Pitch listing”) and copy entire contents of text box
   • If you’ve already got the header line, you don’t need the first line

Data construction: pitch listing

Continued:

3. Paste into spreadsheet
4. In paste menu, split columns using “Custom”; use two spaces ( )

5. Fill in metadata

Times for listings

Unlike before, we need to contextualize the time values and normalize them

• Compare across speakers, who might have different speeds
• Compare across repetitions, because speakers vary their speed

The best way to do this:

1. Count rows in your series and add the total number of samples into each row in a total column
2. Add an index value in a column indicating which number sample we are on (1, 2, 3 … total)
3. Calculate percent of total duration by dividing index / total and adding this to a third column norm
4. Be sure to include –undefined– values as empty cells

Data construction: formant listing

Same as pitch listing, but note there are four columns for F1-F4, plus one for times

Much like above, but with more measurement columns:

Value of listings

Working with listings is more difficult than working with mean measures, but having time series of measurements is valuable

• Limbum question prosody involves boundary tones such as L%
• These exert lowering effects and create phonetic contour tones figure from Gjersøe, Nformi, & Paschen (2019)
• Plots contain average pitch tracks for 300-400 sentence readings, 3 speakers

Coding

Beyond the scope of this tutorial, but more efficient in a number of areas: basic coding

• Praat scripting can more quickly produce tabular data (ask us how!)
• Plots can be made using R (or Excel!)
• Mostly free software

While there is a steeper learning curve, the improvement to the process may pay off in the long run

Acoustics vs. articulation

Acoustics gives us an indirect idea of the movements of the articulators

Sometimes, though, we need to look directly at the articulators

• If there are multiple explanations for acoustics
• If the examined sounds are totally unfamiliar or especially unusual
  • In the Bantoid area, this is not uncommon!

Articulation measurement

We point out here that characterizing and measuring some articulations is much easier than usually expected

• Lip articulation
• Palatography
• Basically photography
  • With whichever resources are on hand
  • All of my personal examples shown here are shot on a phone camera

Lip articulation

Lips are easily seen moving since they are on the exterior of the face

• Mirror at 45 degree angle provides an additional side view of the lips

Example: Babanki lip articulation during vowels Faytak & Akumbu (2021)

• Rounded [u] (right) can be contrasted with labiodental [ʉv] (left) and bilabial compressed [ʉβ] (center)

Model use

Lip activity during Medumba [ʉ]: compressed in the direction of a bilabial stop Olson & Meynadier (2015)

• Relaxed lower lip often leads to bilabial trill [ʙ] during vowel figure from Olson & Meynadier (2015)
• Compression vs. rounding totally unexplored for most Bantoid languages with an /ʉ/

Subtle lip posture differences also known to occur on various fricatives and approximants

• Differences in posture enhance the contrast between labiodental /f, v/ and bilabial fricatives /ɸ, β/ in Ewe, Kwangali, ruGciriku, etc. Ladefoged (1968, 1990); Utman & Blumstein (1994)

Model use

For faster articulations, short videos can also be useful

• Video of a Mono speaker producing a labial flap Olson & Hajek (2004)
• Word is [àⱱétòɾò] “stick used in animal trap” (video repeats)

Palatographs

A “tongue-print” on the roof of the mouth which shows where articulatory activity takes place Anderson (2008)

• Stop and fricative place
• Certain aspects of vowel articulation

The method is more involved than simple photography:

1. Paint tongue with mixture of oil and dark edible powder (chocolate or charcoal)
2. Speaker produces a single token of a word with one lingual consonant
3. Open mouth, insert mirror, photograph

Model use

Kom exhibits a contrast between two vowels which I will discuss further in my regular talk

• Regular [i], as in [ībi᷇] ‘kola nut’
• Fronted, fricativized [i̟], as in [ə̄bí̟] ‘goat’

Palatography reveals aspects of the lingual articulation of these two vowels

• Contrasts like these are probably quite frequent in the Bantoid area
• Compare the [z͡i] vowels observed in Len Mambila and Limbum Fransen (1995); Nforgwei (2004); Connell (2007); Lewis & Shittu (2014)

Model use

More canonically used for clear evidence of lingual place contrasts

• Dental vs. (post)alveolar stop contrast in Lusoga Nabirye, de Schryver, & Verhoeven (2016)

Equipment

Everything that we’ve talked about in this section involves minimal equipment

• Smartphone camera or point-and-shoot camera
• Dental mirror: or hand-held metal mirror, about 6cm wide and 12cm long
  • Held outside for lip angles
  • Inserted against upper molars for palatography
• Edible pigment (chocolate powder, edible charcoal) for palatography
• A brush for painting the tongue (I prefer a narrow paintbrush or a makeup brush)

Optional:

• Phone tripod or camera tripod
• Photography lighting

Figures vs. tabular data

The value of photographic evidence as a figure should be obvious: simply include the image

If you have larger numbers of images, you might convert these to tabular data for further analysis

• Measure the size of some physical attribute and enter that into your spreadsheet
• For example, measure ratio of lip opening width to lip width in pixels
• This normalizes for speaker size and for position (head tilt, here) figure from Olson & Meynadier (2015)

Other articulations

It should be mentioned that certain articulations produced further back are not discussed here

• Dorsal or pharyngeal consonants
• Complex tongue shapes, as in clicks
• Advanced/retracted tongue root

However, ultrasound technology is gradually making it easier to image these articulations Miller, Gick et al. (2006); Namaseb & Iskarous (2007); Allen, Pulleyblank, & Ajíbóyè (2013); Hudu (2014)

• Entirely portable and non-invasive
• But some significant technical barriers remain

File naming and metadata

Name your recording files according to the same logical pattern

• Date, language, topic
• Avoid duplicating names
• Avoid vague names

To identify further details, speak them during the recording itself

• Identify and name yourself
• Identify speaker(s) and their roles if this is not sensitive information
• Identify anyone else who may be heard in the recording (assistants, translators, etc.)
• Give time, date, and location of recording

File backups

Back up every recording in multiple locations if possible, to avoid technological problems or theft destroying your work

• SD cards or thumb drives
• Multiple computers (share key files with a trusted colleague)
• Email small files to yourself (download as attachments later)
• Long-term cloud storage (Google Drive, OneDrive, Box, Dropbox, etc) is ideal but uploading may be expensive

Processing tabular data

Tabular data, of whatever sort, needs to be averaged, summarized (mean/standard deviation), or submitted to a statistical model

• Averaging and some basic statistical models can be done using Excel’s Data Analysis Tools (need to install)
• While Excel and Google Drive (below) work for basic things, learning a dedicated statistical program is better in the long run
• Collaborators may be willing to help

Statistics

Simple statistical analysis and modeling are standard in phonetics for analysis of numerical data (t-testing, linear models, curve fitting, etc)

• We acknowledge this is not practical for all attendees, but collaborators can help guide this work
• Having nicely formatted tabular data is crucial to this handoff

If you want to try stats yourself, it’s best to find software for handling tabular data:

• R is free software which is useful for statistical analysis
• Excel can also be used
• Simple plots can also be made in Excel and R

Statistical power

If you plan to model the data, it’s important to have enough statistical power

• Need to ensure you have collected numerous repetitions of the phenomenon you’re after
• Need to ensure that enough speakers are recorded (the more the better)
  • To get at the community average, instead of the idiosyncrasies of one speaker
• We’ve necessarily used very small data sets for this demo: in reality larger numbers are needed

For the average phonetics study, good power is obtained at 15-20 speakers (rule of thumb)

Summary

We’ve covered:

• Basics of instrumental phonetics
• Praat (annotations, figures, measurements)
• Text and tabular data
• Photography for articulation data
• Practical tips

What remains to be seen:

• Long-term prospects for the methods
• Unanticipated stumbling blocks
• Incorporating further methodological advances

An audit

In compiling these slides and the references they contain (as many papers which present phonetic evidence in African languages as possible), I reflected upon:

• The current state of the phonetics literature for African languages
• Who writes that literature
• Who is cited here

In the references provided here, it is clear that African linguists are underrepresented, regardless of how “African linguist” is defined

• By extraction or by university affiliation
• Regardless of author position (first or other)

Open discussion

Some potential topics for discussion:

• How can we build capacity for this work in Africa’s universities?
• Can existing high-quality documentary work be used as a starting point?
• How can we advise this work when (not “if”) students decide to pursue it?
• What should be the role of linguists off the continent in developing this area?