In 2016, a post-graduate language technology educator specializing in natural language processing and data science published a report titled "Which accents does automatic speech recognition work best for?"
The survey concluded that automatic speech recognition (ASR) performs better for certain English dialects than others.
The researcher analyzed a small sample of YouTube videos and found that, in general, the ASR system worked best for speakers from California—where Google is headquartered—and worst for speakers of Scottish English. Accents that diverged from the general American variety fared even worse.
Although artificial intelligence (AI) has made progress in recognizing accents since then, the diversity of human speech, particularly in terms of accents and dialects, remains a complex challenge for AI.
Studies indicate that ASR systems have higher error rates for some English speakers due to their inability to recognize different phonetic pronunciations and accents.
However, the days of AI missing subtle conversational aspects understood by humans may soon come to an end, as many AI models are now being trained on datasets that include various American phonetics.
Recently, Howard University announced a collaboration with Google Research to release data aimed at improving the automatic speech recognition experience for Black individuals.
This initiative, called Project Elevate Black Voices, involved researchers traveling across the United States to document dialects and speech patterns commonly used in Black communities.
Gloria Washington, Ph.D., a researcher at Howard University and co-principal investigator of the project, emphasized that voice assistant technology should recognize different dialects to effectively serve people who speak them.
In total, researchers collected 600 hours of data representing dialects from 32 states.
Courtney Heldreth, Ph.D., co-principal investigator at Google Research, stated that their mission at Google is to make technology useful and accessible for everyone.
Howard University will maintain ownership of the dataset and oversee its responsible use, ensuring that it benefits Black communities.
Google can also utilize this dataset to enhance its products, making sure its tools work effectively for a broader audience.
Furthermore, researchers at Howard University are actively working to expand training datasets to encompass a wider range of accents and dialects.