Speech AI increasingly mediates access to public services, healthcare, and education, yet it routinely fails speakers of low-resource, Indigenous, and non-standard language varieties. We argue these failures are not merely technical but reflect implicit linguistic policies that reproduce colonial language hierarchies. Drawing on Bourdieu’s linguistic capital, racio-linguistic ideology, and decolonial computing, we show how choices in training data, evaluation metrics, and language model priors operate as political acts that determine whose voices become machine-legible. We introduce the Three Harms (3M) taxonomy - Misrecognition, Misalignment, and Mistrust - to capture failure modes beyond word error rate. We then propose a Participatory Framework for Culturally Competent Speech AI, comprising participatory auditing, community co-design, equitable deployment, and feedback integration, which positions Global South communities as co-designers rather than passive targets of inclusion.
Index Terms: speech recognition, linguistic bias, decolonial AI, low-resource languages, participatory design, cultural competence.