Brink, Jildou (2026) An Investigation of Gender-Based Performance Disparities in Zero-shot WhisperX Mode. Bachelor thesis, Data Science and Society (DSS).
|
PDF
ThesisDSS.pdf Download (504kB) | Preview |
Abstract
Automatic Speech Recognition (ASR) systems are increasingly used in everyday life by communications technologies and accessibility tools. However, there are still concerns related to fairness, and demographic biases especially regarding gender-based differences in transcription accuracy. Existing literature remains inconsistent in its findings. Some studies report higher Word Error Rates (WER) for female speakers, while others show disadvantages for male speakers. Additionally, comparisons between multiple languages remain limited, particularly on lower-resource languages. Therefore, this study investigates whether ASR systems produce varying WER outcomes for female and male speakers across English and Dutch speech data using OpenAI Whisper. Audio samples and reference transcripts from multilingual speech datasets were transcribed using Whisper through the University of Groningen’s Hábrók high-performance computing environment. Word Error Rates were calculated by comparing the generated transcripts with the reference transcripts. To evaluate the effects of gender and language on transcription performance statistical analyses, including a two-way ANOVA test, were conducted. The results showed no statistically significant differences between male and female speakers. Additionally, no significant interaction effect was identified between gender and language. This suggests relatively stable transcription performance across the examined groups. Modern multilingual ASR systems like Whisper may demonstrate improved robustness across genders within high resource language settings. Nevertheless, further research remains necessary to evaluate fairness across more diverse linguistic contexts, accents, and low resource languages such as Frisian.
| Item Type: | Thesis (Bachelor) |
|---|---|
| Name supervisor: | Nayak, S. |
| Date Deposited: | 10 Jun 2026 10:59 |
| Last Modified: | 10 Jun 2026 10:59 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/801 |
Actions (login required)
 |
View Item |