Howard University and Google Research Address AI Speech Bias Against Black Americans
For years, researchers and the public have warned that voice recognition technology favors certain accents and genders. Howard University and Google Research are addressing this through the Project Elevate Black Voices partnership, according to an article in Howard University’s The Dig.
African American Vernacular English (AAVE) or Ebonics is a language rooted in history and culture. Inherent bias in the development of voice recognition technology often produces incorrect results when Black users vocalize commands. Many Black users have been conditioned to change their voice patterns (code-switch) away from their natural accents in order to be understood by voice recognition products.
Howard and Google researchers traveled around the country collecting more than 600 hours of different AAVE dialects from 32 states. They catalogued the dialects and diction commonly used in Black communities that are misunderstood or not always recognized by AI technology.
The researchers recently released the dataset in the hopes that it will improve automatic speech recognition technology. AI developers train large language models (LLMs) using such datasets, making it critical to have access to diverse accents, dialects and languages.
Howard University will retain ownership of the dataset and licensing, and serve as steward for its responsible use.