Why do speech-to-text systems fail more often for Black speakers?

Koenecke and colleagues tested five commercial speech-to-text systems — Amazon, Apple, Google, IBM, and Microsoft — on 19.8 hours of sociolinguistic interview audio matched by age and gender across white and black speakers, and found "substantial racial disparities" in every system: an aggregate word error rate (WER) of 0.35 for black speakers against 0.19 for white speakers. The gap held system by system — Microsoft's best-performing ASR still produced a WER of 0.27 for black speakers versus 0.15 for white speakers, and Apple's worst-performing ASR produced 0.45 versus 0.23 — "the error rates for black speakers are nearly twice as large in every case."

The authors trace the disparity to the acoustic model rather than to vocabulary or language-model effects: "the race gap was equally large on a subset of identical phrases spoken by black and white individuals," which rules out word choice as the driver. Their matched corpora — the Corpus of Regional African American Language (CORAAL, speakers of African American Vernacular English across three US sites) against Voices of California — isolate pronunciation and prosody as the source of the error, and the paper's proposed fix is "using more diverse training datasets that include African American Vernacular English."

This sits in tension with Do speech models learn language-specific sounds or universal physics?, which credits self-supervised speech models with learning universal, dialect-agnostic articulatory representations that transfer across languages. Koenecke et al.'s commercial acoustic models show the opposite inside a single language: WER nearly doubles for a dialect (AAVE) the training data underrepresented, so whatever acoustic universality SSL models achieve across languages does not automatically extend across dialects within one. It also sharpens What speech tasks remain without standardized benchmarks?: even the narrow transcription-quality metric that note worries evaluation has overfit to is itself unevenly accurate across demographic groups, so narrowing evaluation to transcription doesn't even guarantee a fair measure of transcription.

The study measures five 2019-era commercial systems on two regional US English corpora; it does not test non-English ASR, more recent end-to-end or LLM-based speech systems, or dialects beyond AAVE, so the specific error-rate gap should not be read as a permanent or universal property of ASR generally. What it does establish is that a widely deployed class of consumer tools — virtual assistants, captioning, dictation — was, at the time of testing, roughly twice as error-prone for black speakers as for white speakers, traced to training-data composition rather than isolated edge cases, which licenses auditing any voice-driven product for demographic skew before calling it inclusive.