Today

1. A mock speaker comparison
2. Profiling an unknown speaker
3. Disguise
4. What the class data show

Groups and folders

• You work in groups of three. Each group receives a case folder, a profiling recording and a set of disguise recordings. File names in the folders are codes, and do not show whose voice it is
• Keep the measurements from the last session at hand: you need the same procedures today

The case

Your folder contains:

• one questioned recording, Q.wav: sentences B, read by one speaker
• four known recordings, K1.wav to K4.wav: sentences A, read by four “suspects”

Step 1: listen first

• Listen to Q several times, then to each K. Write down your auditory impressions before you measure anything: accent, voice quality, speed, anything unusual. Listen in particular to the /r/ in dark, water, car and heard, and the /t/ in water
• Note whether you think you recognise a classmate, and at what point. A familiar voice can bias the rest of the analysis (sessions 5 and 10)
• Keep the order in which you listened, and decide in the group how to avoid reading your first impression into the measurements

Step 2: measure

For Q and each K:

1. Median f0 and standard deviation (Pulses > Voice report)
2. Articulation rate: 42 syllables in Q, 41 in each K
3. F1 and F2 at the midpoint of the stressed vowels in dark, water and heard, which occur in both sets

Step 3: similarity and typicality

For each feature, ask how close Q is to each K, its similarity, and how common Q’s values are in the relevant population, their typicality. Here, the population is the class, from the spreadsheet of the last session.

• A shared value that half the class also has is weak evidence. A shared value that is rare in the class is stronger

The class as a population

Simulated preview. Q and K2 are similar, but several classmates have a median f0 in the same range, so f0 alone does not set K2 apart. In class we look at the real distribution in R.

Step 4: conclusions

• For each suspect, combine the features into one conclusion on the verbal scale of session 4, from strong support for different speakers to strong support for the same speaker. Give the two or three observations that weigh most, and say which differences you think are within-speaker variation
• Present your conclusions to the class before the answers are revealed

Discussion

Two members of your group reach different conclusions about the same suspect.

Should the report give the majority view, the more cautious one, or both? How might a court read each option?

An unknown speaker

Your group receives the word list and sentences A of a classmate who is not in your case folder. Write a short profile (session 3):

• first language and, if you can, region, and other languages that might have influenced the speaker’s English
• anything else the voice suggests, such as an age range

Checking the profile

• Your lecturer checks your profile against the information on the speaker’s form
• Which claims were right, which features led you there, and which claims went beyond what the recording could show?

Who is behind the disguise?

Your group receives four disguised readings of sentences A (D1–D4) and four normal readings (N1–N4) by the same four speakers, in shuffled order.

1. Match each disguised reading to a normal one, by ear first
2. Then measure median f0, its standard deviation and articulation rate in each recording, and check whether the numbers support your matches
3. Write down how confident you are in each match

What did the disguise change?

For each pair you matched:

• Which strategy do you think the speaker used: pitch, accent, voice quality, speed?
• Which features changed, and which stayed the same?
• Which features helped you see through the disguise?

Disguise in the class

Simulated preview. In class, we plot your normal and disguised readings, with the strategy from your forms.

Discussion

• How did knowing some of the voices affect your analysis, and how could a lab prevent this?
• Which features were most useful in the comparison, and which varied most between sentences A and B?
• How far did the class data work as a population? Who would be the relevant population in a real case?

Summary

Slides and code

These slides

mcanzi.de/ws27_fss/12_praat_casework.html

Code and all sessions

github.com/mcanzi/ws27_fss