Listen and Speak Fairly: A Study on Semantic Gender Bias in Speech Integrated Large Language Models

TL;DR

This paper investigates semantic gender bias in Speech Integrated Large Language Models across multiple tasks, revealing language-dependent bias variations and emphasizing the need for diverse evaluation methods to ensure fairness.

Contribution

Introduces a curated bias evaluation toolkit and dataset for assessing gender bias in SILLMs across four semantic tasks, highlighting bias variability and evaluation challenges.

Findings

Bias levels vary with language and task

Multiple evaluation methods are necessary for comprehensive bias assessment

Bias can amplify existing societal stereotypes in speech models

Abstract

Speech Integrated Large Language Models (SILLMs) combine large language models with speech perception to perform diverse tasks, such as emotion recognition to speaker verification, demonstrating universal audio understanding capability. However, these models may amplify biases present in training data, potentially leading to biased access to information for marginalized groups. This work introduces a curated spoken bias evaluation toolkit and corresponding dataset. We evaluate gender bias in SILLMs across four semantic-related tasks: speech-to-text translation (STT), spoken coreference resolution (SCR), spoken sentence continuation (SSC), and spoken question answering (SQA). Our analysis reveals that bias levels are language-dependent and vary with different evaluation methods. Our findings emphasize the necessity of employing multiple approaches to comprehensively assess biases in SILLMs, providing insights for developing fairer SILLM systems.