Covert racism: the voice inside the machine

Speech is one of humanity's earliest interfaces. Writing is a later technology layered on top of it. Yet much of contemporary AI has inverted that history, treating voice as though it were simply text delivered in audio form.

A speech system reduces what you said to generic text at the earliest possible stage, and the rest of the signal — timing, emphasis, self-repair, overlap, dialect — is discarded as noise. In low-stakes use that is a reasonable trade. Your assistant mishears you, you repeat yourself, you move on.

People under pressure, with unequal power, speaking in linguistically diverse ways, convey meaning through pauses, emphasis and dialect rather than vocabulary alone. A system that collapses that into normalized text has quietly installed one way of speaking as the default and every other as an error. What follows is not a usability complaint. It is an allocational harm — who gets opportunity, care and trust.

The clearest evidence is the oldest and least contested: commercial speech recognition is substantially worse at Black speakers, and the gap is not about vocabulary.

Transcription error is the visible failure, and the one a vendor can be held to. The harder finding is what happens downstream, in the model that reads the transcript.

Hofmann and colleagues found that language models hold covert prejudice against African American English — assigning less-prestigious jobs, and convicting more often, on the basis of dialect alone, in scenarios where race is never mentioned and the statements have nothing to do with the decision.

Put the two halves together. A transcription layer that hears one group less accurately hands a degraded record to a decision layer that already treats dialect as a signal about the person. Neither layer announces what it did. Both are, individually, defensible engineering choices.

This is the argument that produced everything else we build. It is why prosody is never a score in our own assessment, why accent and pace and hesitation never reach an employer as a judgment, and why a typed path is not an accommodation bolted on afterwards but the same path, chosen at the start.

It is also the reason we treat measurement rather than screening as the deliverable. You cannot filter out a bias whose defining property is that it does not show up when you ask about it. You can only instrument the system so the disparity becomes visible, reproducible and arguable — and then argue about it.