Language technology practitioners as language managers: arbitrating data bias and predictive bias in ASR

ASR systems systematically fail for non-standard language varieties: Automatic speech recognition performs significantly worse on marginalized and non-standardized language varieties compared to standardized ones, despite language variation being a natural characteristic of human speech.

Data bias in training creates predictable algorithmic harm: The paper identifies that current industry practices in how ASR systems are trained and tested create systematic data bias—meaning certain groups of speakers experience reliably worse performance, which is not a random technical glitch but a predictable outcome of design choices.

Language policy frameworks reveal hidden choices: By analyzing ASR development through a language policy lens, the authors show that practitioners are making implicit policy decisions about which language varieties "count" and deserve investment, similar to how governments make official language decisions—these choices are often invisible but have real consequences for whose voices the technology serves.

Language resources should be treated as public infrastructure, not just commercial products: The paper argues that datasets and language resources used to build speech technology should be designed with input from affected speech communities themselves, rather than solely optimized for market viability, similar to how public infrastructure like roads serve broad populations rather than just profitable segments.

Practitioners have agency in mitigating bias: The framing suggests that language technology professionals can actively reduce algorithmic bias by understanding these policy dimensions and making intentional choices about whose language varieties are represented in training data and testing procedures.