Lab Report

DEMOGRAPHIC FACTORS IN THE ACCURATE VOCAL AGE RECOGNITION OF ASHKENAZI JEWISH WOMEN

Ben Fink and Ari Odinec

June 5, 2025

Intensive Biology

Abstract

                The purpose of this experiment is to discern if ethnicity and/or gender play a role in age estimation based on voice. The first hypothesis states that if the gender of a listener is the same as the gender of a speaker, the listener will be able to guess the speaker's age more accurately than other genders. The second states that if the ethnicity of a listener is the same as the ethnicity of a speaker, the listener will be able to guess the speaker's age more accurately compared to other ethnicities. 91 9th graders were asked to anonymously estimate the age groups of 10 Ashkenazi Jewish female voices. The 34 Ashkenazi Jewish listeners averaged guessing 8/10 Ashkenazi Jews’ ages, which was (according to the t-tests) significantly more accurate than both the 57 non-Ashkenazi Jews’ average of 7/10 and the whole group’s average of 7.37/10. On the contrary, the 44 males who averaged about 7/10 correct age-group estimations and the 47 females who averaged about 7.5/10 correct estimations were not significantly different from each other or the whole group (according to the t-test). Thus, the hypothesis about ethnicity is confirmed, supporting that Ashkenazi Jews recognize the ages of Ashkenazi Jews based solely on voice more accurately than other ethnicities. And the portion of the hypothesis on gender was refuted, supporting that gender/having the same gender does not impact the accuracy of age estimation based on voice.

Background Information/Introduction

Our experiment presented the voices of 10 Ashkenazi Jewish women of varying age groups to 91 9th graders (14-15 years old), who were then asked to guess the age-groups of the voices. The gender and ethnicity of the respondents were anonymously collected. We then measured the mean accuracy rate out of 10 for the whole group to use as a baseline and the mean accuracy rate for the specific identity groups. Our objective was, through these comparisons, to address whether the speaker and listener having the same gender and/or ethnicity will increase the listener's ability to guess the age of the speaker.

Prior studies suggest that age can affect many factors related to voice. An article from Weill Cornell Medicine shows that aging causes atrophy and posture changes that can make it more challenging to make the same sounds as when a person was younger.^[1] Aging can cause higher pitch of voice in women, lower in men, voice tremors, and more.^[2] However, the effect of a listener’s ethnicity and gender on their ability to notice these nuances has not been extensively researched. It has been established that both gender and ethnicity influence age estimation from faces, but the influence on age estimation from voices has not clearly been tested.^[3]

Variables & Hypothesis

Hypothesis: If the gender of the listener is the same as the gender of the speaker, the listener will be able to more accurately guess the speaker's age compared to other genders.

Reasoning: Previously established own-group biases like language and age support the conjecture that spending more time with a group (usually resulting from belonging to that group) increases age estimation based on voice.^[4] This can be applied to gender, as socially, females tend to spend more time with females and less with males, and vice versa.^[5]

Hypothesis: If the ethnicity of the listener is the same as the ethnicity of the speaker, the listener will be able to more accurately guess the speaker's age compared to other ethnicities.

Reasoning: Own-ethnicity bias with age estimation based on faces has been established in prior studies.^[6] We hypothesize that the same familiarity with one's ethnicity which improves age estimation from faces will also increase accuracy in estimating age from voices.

Independent variable: Ethnicity of the listener and Gender of the listener

Dependent variable: Average number of correct age group identifications (out of ten speakers) based on voice alone (mean questions correct out of 10) for each ethnicity group and each age group

Materials/ Methods

1. We enlisted 10 Ashkenazi Jewish and female volunteers; there are two volunteers who belong to each age group (Child: 6-11, Adolescent: 12-18, Young Adult: 19-35, Middle-Aged: 36-64, Senior: 65+).  (This group will be referred to as “speakers.”)
2. Each speaker was sent a Google Form that asked them to specify their name and age, confirm their ethnicity and gender, and submit a recording of them saying, “A king ruled the state in the early days. The young prince became heir to the throne.”
3. Then, a form was created which we attached these 10 recordings to with anonymous names and no other information provided to be sent out to respondents. The form asks for the respondent’s gender and ethnicity but not name or email.
4. 91 9th grade students completed the form. Each listener should listen independently to each full recording before the form allows them to select the age group, and they will not be told any information about the speakers.
5. The number of correct age-group estimates made by each listener out of 10 total speakers was recorded. The accuracies of the whole groups were averaged using the arithmetic mean. The results of Ashkenazi and non-Ashkenazi groups would also be averaged, as well as with female and male groups.

*This experiment was conducted once. (One repetition.)

Figure 1: Menu Used to Collect Subject Information on Respondent Form

This image shows the menu used on the form sent to respondents to collect gender and ethnicity. Respondents were required to fill out this section before moving forward.

Figure 2: Menu Used to Collect Respondent Age Guesses on Respondent Form

This image shows the menu used on the form to guess the age of the speaker. It would appear after the blue “play” button was pressed and the recording was played through fully to ensure that the respondent actually listened to the recording.

The choices made for this experiment were all based on scientific reasoning to further the experiment’s reliability. Ashkenazi (Eastern European) Jewish women were chosen as the control group to test if a person is more accurate at guessing the age of someone of their own gender or ethnicity without seeing the person and only hearing their voice. We chose 9th graders as the listeners because they all share the same age, and we had over 90 willing volunteers. We chose “A king ruled the state in the early days. The young prince became heir to the throne” because it was a combination of two Harvard sentences, which are used to test intelligibility of telecommunications systems.^[7] The age groups that the experiments were based upon are as shown: Child: 6-11, Adolescent: 12-18, Young Adult: 19-35, Middle-Aged: 36-64, Senior: 65+. They were chosen based on an article with various data put together by Lumen Lifespan Development, though we chose to combine the last two ages into 65+ because the elderly are categorized as 65+ by the American government.^[8] We used a Google Form to collect recordings and directed speakers to record in quiet settings to ensure reliability and uniformity. We chose to design a form from scratch to send out to respondents because it allowed us to build error prevention into our collection process.

Results

Writing

Figure 3 shows the mean accuracy of the listener’s age group estimations (out of ten), specifically looking at the ethnicity groups.

• Whole Group: 7.37/10 (mean) questions correct, ± 1.41 (AAD).
• Ashkenazi Jews: 7.97/10 (mean) questions correct, ± 1.04 (AAD).
• non-Ashkenazi: 7.02/10 (mean) questions correct, ± 1.46 (AAD).

Figure 4 shows the mean accuracy of the listener’s age group estimations (out of ten), specifically looking at the gender groups.

• Whole Group: 7.37/10 (mean) questions correct, ± 1.41 (AAD).
• Males: 6.95/10 (mean) questions correct, ± 1.29 (AAD).
• Females: 7.55/10 (mean) questions correct, ± 1.32 (AAD).

Figure 5 shows the results of the T-tests performed on combinations of the different ethnicity categories and the different gender categories.

• Table 3

• P-value for the whole group compared to males: 0.19; we cannot reject the null hypothesis.
• P-value for the whole group compared to females: 0.56; we cannot reject the null hypothesis.
• P-value for males to females: 0.10; we cannot reject the null hypothesis.

• Table 4

• P-value for the whole group compared to Ashkenazi: 0.049; we can reject the null hypothesis.
• P-value for the whole group compared to non-Ashkenazi: 0.24; we cannot reject the null hypothesis.
• P-value for non-Ashkenazi vs. Ashkenazi: 0.01; we can reject the null hypothesis.

Figures

Figure 3: Mean Accuracy of Age Estimation Based on Voice Full Group As Well As Ethnic Categories

Respondents of different ethnicities and genders were asked to listen to 10 different Ashkenazi female voices and guess their ages. Results show their mean accuracy out of ten questions for each ethnic group. Column 1 shows mean questions correct out of 10, column 2 shows mean percent questions correct, column 3 shows average absolute deviation, and column 4 shows percent average absolute deviation. 91 total respondents, one repetition.

Respondents of different ethnicities and genders were asked to listen to 10 different Ashkenazi female voices and guess their ages. Results show their mean accuracy out of ten questions for each ethnic group. Blue represents the mean for all ethnicities, red represents Ashkenazi respondents, and yellow represents non-Ashkenazi respondents. The x-axis represents ethnicity, and the y-axis represents the mean number of questions correct out of 10. Error bars represent ± AAD, 91 total respondents, one repetition.

Figure 4: Mean Accuracy of Age Estimation Based on Voice Full Group As Well As Gender Categories

Respondents of different ethnicities and genders were asked to listen to 10 different Ashkenazi female voices and guess their ages. Results show their mean accuracy out of ten questions for each gender group. Column 1 shows mean questions correct out of 10, column 2 shows mean percent questions correct, column 3 shows average absolute deviation, and column 4 shows percent average absolute deviation. 91 total respondents, one repetition.

Respondents of different ethnicities and genders were asked to listen to 10 different Ashkenazi female voices and guess their ages. Results show their mean accuracy out of ten questions for each gender group. Blue represents the mean for all genders, red represents male respondents, and yellow represents female respondents. The x-axis represents gender, and the y-axis represents the mean number of questions correct out of 10. Error bars represent ± AAD, 91 total respondents, one repetition.

Figure 5: T-Tests

Female and Male respondents were asked to listen to 10 female voices and guess their ages. The table shows T-tests performed on different combinations of the included gender groups shown (as labeled in each column) P-values all above 0.05, thus cannot reject the null hypothesis; there is likely no significant correlation between a person’s gender and their ability to guess the age of a voice of the same gender. 91 total respondents, one repetition.

Respondents of different ethnicities were asked to listen to 10 different Ashkenazi female voices and guess their ages. The table shows T-tests comparing the various groups (each column indicates a different comparison). P-values comparing Ashkenazi vs. whole group and Ashkenazi vs. non-Ashkenazi group both below 0.05, thus rejecting the null hypothesis for those comparisons; the difference is likely statistically significant. 91 total respondents, 10 speakers, one repetition.

Data Analysis

Hypothesis:

• Having the same ethnicity as the speaker will allow the listener to recognize the age of the voice more accurately.

Evidence:

• Ashkenazi Jewish respondents averaged guessing 7.97 ages correctly of the 10 speakers which is almost 10% more than non-Ashkenazi respondents’ 7.02 correct questions right
• T-tests comparing the Ashkenazi group to the total group as well as to the non-Ashkenazi group showed statistical significance in the results; the p-values were under 0.05 allowing us to refute the null hypothesis.

Conclusion:

• Ashkenazi listeners outperformed the rest of the respondents across the board
• Confirmed hypothesis: If the ethnicity of the listener is the same as the ethnicity of the speaker, the listener will be able to more accurately guess the speaker's age compared to other ethnicities.

Hypothesis:

• Having the same gender as the speaker will allow the listener to recognize the age of the voice more accurately.

Evidence:

• Females averaged guessing 7.55 age groups correctly out of the 10 speakers, and males averaged guessing 6.95 showing a difference of only 0.6 questions on average per person.
• The t-tests comparing female data to male data and female data to the general population were both over .05, meaning we cannot reject the null hypothesis

Conclusion:

• There is no measurable difference between males and females with age guessing accuracy based on voice.
• Refuted hypothesis: The gender of the listener being the same as the genders of the speaker will not impact the accuracy at which they estimate ages of voices.

Previous Data

The Effect of Language Familiarity on Age Perception- Nagaou and Kewley Port

• Findings:

• “Age estimation was more accurate for familiar language than less familiar language.”^[9]

• Correlation with our findings:

• Both support the idea that familiarity and exposure either to a language or to an ethnicity causes a more accurate age estimation

*There were no studies comparable to the gender portion of this experiment and the study on language was the best comparison for ethnicity (shown above)

Significance of your findings

• The concept that familiarity with an ethnicity leads to higher auditory perception can help to make advancements in cross-cultural communication and reduce misunderstandings.^[10]
• Implications for Forensic and AI Applications: Understanding the effect of ethnicity can be used to help investigative police activity and allow them to create accurate age estimation AI systems that account for ethnicity.^[11]
• Diverse Data in Technology: The findings show how important it is to include a broad spectrum of ethnic groups to create datasets for voice recognition so technologies ensure equal accuracy across all ethnicities.^[12]

Possible Sources of Error

• We did our best to avoid all systematic errors by creating a form from scratch, but there could have been some errors that could not have been avoided like…

• Students could have shared their answers to the form with each other (random error, over or underestimation)
• Students could have misreported their own ethnicity or gender (random error, over or under estimation)

• Students may have recognized some of the people (random error, increase in overall accuracy for all groups)
• Computer audio quality was not the same across the board-could have caused voices to be heard differently depending on the computer (random, increase or decrease for all data)
• Though ethnicity and gender were controlled, other variables could have affected voice which were not controlled (socioeconomic, place of origin, etc) and the relationship with those identifiers and the listeners and speakers were not monitored (random, increase or decrease in data)

Future Experiments

• To test the versatility of our results, we could do the same experiment but instead of Ashkenazi Jewish, we would test multiple other ethnicities.
• Testing if having the same accent or growing up around a certain accent causes higher accuracy in guessing someone's age based on their voice
• Testing if having the same ethnicity or gender causes a higher accuracy in guessing someone's age based on their appearance

Conclusions/Future Directions

Ashkenazi Jews recognize the ages of Ashkenazi Jews based solely on voice more accurately than other ethnicities

Gender/having the same gender does not impact the accuracy of age estimation based on voice.

Bibliography

"Aging Voice." UT Southwestern Medical Center. Accessed May 22, 2025. https://utswmed.org/conditions-treatments/aging-voice/.

Blakemore, Erin. "Here's How Your Voice Changes as You Age." Sean Parker Institute for the Voice. Weill Cornell Medicine. Last modified June 4, 2024. Accessed May 21, 2025. https://voice.weill.cornell.edu/about-us/news-and-updates/here%E2%80%99s-how-your-voice-changes-you-age.

"Harvard Sentences." Columbia University. Accessed May 22, 2025. https://www.cs.columbia.edu/~hgs/audio/harvard.html.

Hutiri, Wiebke Toussaint, and Aaron Ding. Bias in Automated Speaker Recognition. https://doi.org/10.48550/ARXIV.2201.09486.

Moyse, Evelyne. "Age Estimation from Faces and Voices: A Review." Psychologica Belgica 54, no. 3 (2014): 255-65. https://doi.org/10.5334/pb.aq.

Nagao, Kyoko, and Diane Kewley-Port. "The Effect of Language Familiarity on Age Perception." Pennsylvania State University. Last modified October 10, 2005. Accessed May 22, 2025. https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=e282bc20e0f76bfba0bf1e8fc7165ea8236ce7ff.

Onnela, Jukka-Pekka, Benjamin N. Waber, Alex, Pentland, Sebastian Schnorf, and David Lazer. 2014. "Using Sociometers to Quantify Social Interaction Patterns." arXiv. doi:10.48550/ARXIV.1405.6224.

Paver, Alice, David Wright, Natalie Braber, and Nikolas Pautz. "Stereotyped Accent Judgements in Forensic Contexts: Listener Perceptions of Social Traits and Types of Behaviour." Frontiers in Communication 9 (January 17, 2025). https://doi.org/10.3389/fcomm.2024.1462013.

"Periods of Human Development." Lumen Learning. Accessed May 21, 2025. https://courses.lumenlearning.com/wm-lifespandevelopment/chapter/periods-of-human-development/.