Howard University And Google Team Up To Advance AI Speech Recognition For African American English

In a significant move toward advancing inclusivity in technology, Howard University and Google Research have unveiled a new dataset designed to enhance how automatic speech recognition (ASR) systems serve Black users. The collaboration, part of Project Elevate Black Voices, involved researchers traveling nationwide to document the unique dialects, accents, and speech patterns commonly found in Black communities, features often misinterpreted or ignored by current AI systems.

The project spotlights African American English (AAE)—also known as African American Vernacular English, Black English, Ebonics, or simply “Black talk”—a culturally rich and historically rooted linguistic form. Due to systemic bias in the development of AI tools, Black users have frequently encountered errors or been misunderstood by voice technologies, sometimes feeling pressured to alter their natural speech just to be recognized by these systems— a classic form of code switching.

Researchers at Howard University and Google are on a mission to change this.

“African American English has been at the forefront of United States culture since almost the beginning of the country,” shared Gloria Washington, Ph.D., a Howard University researcher and the co-principal investigator of Project Elevate Black Voices, in a press release. “Voice assistant technology should understand different dialects of all African American English to truly serve not just African Americans, but other persons who speak these unique dialects. It’s about time that we provide the best experience for all users of these technologies.”  

To build this groundbreaking dataset, researchers gathered 600 hours of speech from participants representing various AAE dialects across 32 states. The goal was to confront hidden barriers that hinder the effectiveness of automatic speech recognition (ASR) systems for Black users. One of the key findings was that AAE is significantly underrepresented in existing speech datasets, not because the language isn’t spoken, but because many Black users have been socially conditioned to alter their natural speech when interacting with voice technology. This phenomenon, often rooted in the need to be understood by systems that don’t recognize AAE, leads to a lack of authentic representation.

A 2023 Google blog post highlighted another challenge: privacy and security policies, while essential, create additional constraints on the collection of AAE-specific voice data. These self-imposed limits make it harder to amass the scale and authenticity of data required to close the gap in performance.

Despite these challenges, progress is being made. Researchers are now using dialect classifiers to identify AAE within broader datasets, a promising first step toward building more inclusive technologies. Howard University will maintain ownership and licensing rights to the dataset, serving as its ethical guardian to ensure it’s used responsibly and for the benefit of Black communities. Google, in turn, will be able to use the dataset to enhance its own ASR products, part of a broader effort to make AI tools more equitable across dialects, languages, and accents globally.

SEE MORE:

What Are Racial Microaggressions?

Black Culture, White Face: How the Internet Helped Hijack Our Culture

Read more