Decolonizing Automatic Speech Recognition Policies

Summary & Methodology Analysis

The paper addresses how current automatic speech recognition systems systematically fail speakers of low-resource, Indigenous, and non-standard language varieties because they function as linguistic policies that reproduce colonial language hierarchies. To tackle this, the authors develop a theoretical synthesis of linguistic capital theory, raciolinguistic ideology, language policy research, and decolonial computing to define structural linguistic policies in automatic speech recognition. They introduce the Three Harms taxonomy covering Misrecognition, Misalignment, and Mistrust, alongside a seven-layer situatedness model to map linguistic diversity factors like nation, ethno-linguistic identity, and ideology. Furthermore, they define a minimum audit protocol that includes defining context and risk, constructing culturally grounded test sets, specifying diverse evaluator roles, annotating failure modes, and conducting community-led adjudication. They also propose a four-pillar participatory framework consisting of participatory auditing, community co-design, equitable deployment, and feedback integration, implemented through a continuous loop of repair and redress where community stakeholders exercise authority over evaluation, design, and governance decisions.

From a performance and metric perspective, the paper demonstrates that standardized metrics like Word Error Rate fail to capture harms for tonal languages, whereas Tone Error Rate helps reveal meaning-changing errors that Word Error Rate collapses. Systems consistently demonstrate performance gaps, such as achieving 5 percent Word Error Rate for Standard American English compared to 35 percent Word Error Rate for African American Language. The research draws on datasets and models including UGSpeechData, IndicSUPERB, Kathbath, Mozilla Common Voice, and Siri, though the paper does not specify precise hardware requirements, latency numbers, token counts, or training costs for these models.

Despite the comprehensive framework, the authors note clear limitations. The proposed framework is broad and requires contextual adaptation for each application, as the four-pillar structure is a simplification of complex real-world processes. Additionally, small language communities may lack the resources to support large evaluator panels required for the full audit protocol. The paper does not specify alternative mitigation strategies for resource-constrained communities beyond noting this limitation.