Howard University and Google Research Collaborate to Expand A.I. Recognition of African American English

Artificial intelligence has evolved into a common part of everyday life, from voice assistants to speech-to-text tools. But when technology struggles to recognize the way Black people of varied regions speak, using these tools becomes another exercise in code-switching. Howard University is working with Google Research to change that by teaching artificial intelligence to better recognize the diversity of African American English.

Through Project Elevate Black Voices, researchers from the historically Black university and Google Research collected 600 hours of audio from speakers across 32 states to develop a dataset that reflects the different dialects, diction and speech patterns found throughout Black communities. The goal is to help automatic speech recognition (ASR) technology, which converts spoken language into text or commands, more accurately understand Black users without requiring them to alter their natural speech — or in other terms, code switch.

African American English, also known as African American Vernacular English, or Ebonics, is a language variety with deep cultural and historical roots. Yet, speech recognition systems have often struggled to interpret its nuances. Researchers found that Black users have been conditioned to modify their speech when interacting with these technologies, limiting the availability of natural African American English in the data used to train them. Code-switching, or adjusting our language depending on the social setting, can further complicate the process.