Accent = Worth? Report
Accent Sub-Cultures, Perceived Worth and Social Bias in Punjab, Pakistan: Project Report with Research Survey Results
1. Project Overview
This project aimed to investigate the socio-psychological effects of accent subcultures in Punjab, Pakistan, and how they relate to perceived worthiness, intelligence, employability, social status, and trustworthiness. By pairing acoustic (sound) data analysis with perception surveys, the study empirically tests whether measurable phonetic features correlate with social bias. Its wider mission is to contribute to conversations on linguistic discrimination in South Asia, an under-researched context in the global accent-bias literature.
2. Methodology
Research and design: Conducted a thorough literature review of global accent-bias studies and the sociolinguistics of South Asia and Punjab; formed an ethical framework (consent forms, anonymisation)
Data collection: Recordings of 40 to 50 individuals answering questions about personal choices, such as their favourite season and month. Online perception surveys, in which listeners rate recorded voices on traits such as intelligence and perceived rank.
Analysis: Use of acoustic features (formants, pitch, intensity, vowel space) extracted with Praat and Python; machine-learning techniques (clustering and classification) used to explore accent groups; statistical analysis (ANOVA, regression) correlating acoustic features with survey ratings; and qualitative analysis of interview themes.
3. Survey Results
The online survey received 51 responses. Respondents listened to four audio samples and, for each, answered six questions: perceived education level, likelihood of hiring the speaker for an office job, perceived social class, perceived profession, intelligence rating, and likelihood of trusting the speaker in a high-ranking managerial position. Table 1 shows the most common response per sample, and Figure 1 compares the positive ratings.
4. Key Findings
Clear split in perceived status: Samples 2 and 3 were placed in the upper-middle class by about half of respondents (53% and 51%) and were mostly seen as office professionals, whereas Sample 4 was placed in the lower-middle class (55%) and identified as a blue-collar worker by 84%.
Education is inferred from the voice: About half of the respondents credited Samples 2 and 3 with undergraduate-level education, while 49% believed the Sample 4 speaker had no formal education at all. Sample 1 fell in between, most often placed at Matric/O-Level.
Employability follows the same order: Roughly 53% to 59% rated themselves likely or very likely to hire Samples 2 and 3, compared with about 29% for Sample 1 and 12% for Sample 4. For Sample 4, 65% were unlikely or very unlikely to hire.
Trust in leadership is lowest for Sample 4: Only about 10% would trust the Sample 4 speaker in a senior managerial role, and 37% said “very unlikely”. Samples 2 and 3 were rated likely by 49% and 43%.
Intelligence ratings mirror status: Above-average ratings were about 65% for Sample 2 and 49% for Sample 3, versus 14% for Sample 4, though Sample 4 was still most often rated “average” (57%), not below average.
Sample 1 is the ambiguous case: It drew mostly neutral or average judgements (51% neutral on hiring, 53% average on intelligence), and its perceived class was split between middle and lower-middle.
5. Discussion
Across all six measures, the four samples rank in a consistent order (Samples 2 and 3 highest, then Sample 1, then Sample 4). This pattern is what the accent = worth hypothesis predicts: listeners appear to convert a few seconds of speech into judgements about education, class, occupation, competence, and trustworthiness. Because respondents heard only audio, these judgements are based on how the speakers sound rather than on any other information about them.
6. Next Steps
After collecting accent samples, we used Parselmouth in Python to extract the acoustic features of each recording. There were about ~40 features in total, which included pitch, intensity, jitter, voicing fraction, and harmonicity. Next, we used the data recorded from our survey of employability perception to create an aggregate "employability score" on a scale of 1-5. We normalized qualitative answers such as "Likely" and "Very high" to qualitative values, averaging them for the final score. We trained a Random Forest on the accent samples that we surveyed to predict the employability score for the rest of the recordings. Furthermore, we used a K-means clustering algorithm (optimized at 4 clusters via silhouette scores) to discover natural acoustic groupings in the data. These clusters allowed us to visualize the dataset and analyze how different acoustic profiles correlated with the final predicted scores.
Below are figures that visualize these workings:
