Whose Voice Speaks Volumes? The Problem with Gender Identification from Speech

This study investigates the topic of voice-based gender identification, focusing on aspects that concern accuracy and ethical implications. One aim is to review the literature available in order to understand what shapes the expectations of gender as revealed through voice. Then, an experimental investigation is conducted to examine the impact of age on voicebased gender classification. Our evaluation of a Multilayer Perceptron neural network model using MFCC features indicates 0.90 overall accuracy in binary gender classification. This seemingly high accuracy score would mask a substantial 10% misclassification rate in real-world applications. When considering age, accuracy drops further, with varying scores across the seven groups (0.64-0.88). Generally, our results indicate age-related variability in model performance, highlighting limitations and ethical concerns in generalizing across age groups. As concerns societal implications, our literature review and experiments suggest that greater awareness and diversity-informed approaches are needed in the design, development, and marketing of speech technologies.