topic Scenario: Racial Disparities in Automated Speech Recognition in Trustworthy AI

Scenario: Racial Disparities in Automated Speech Recognition

CatTruxillo — Fri, 02 Feb 2024 21:20:40 GMT

This is a discussion forum for the activities in the Inclusivity module of the Free SAS e-learning course, Responsible Innovation and Trustworthy AI.

Scenario: Racial Disparities in Automated Speech Recognition

Consider This:

What other considerations should be made to ensure that ASR systems are more inclusive?

Re: Scenario: Racial Disparities in Automated Speech Recognition

Rogerio_Neiva — Fri, 04 Oct 2024 14:23:05 GMT

Citing the text "Mitigating the effects of these systems requires thorough auditing and including diverse audio training data sets with African American Vernacular English."

But, can't we say the same about how people talk differently in other parts of the US ? Why should an African American speak differently ? What about Asians ? Latinos ?

And, who chose the speakers ? From where ?

And a last question ? Isn't there an initial bias in the way the test was done to lead to a predefined result ?

Re: Scenario: Racial Disparities in Automated Speech Recognition

jomana-khatib — Fri, 01 Aug 2025 12:10:56 GMT

To make ASR (Automated Speech Recognition) systems more inclusive, we must go beyond diverse training data. It's essential to include dialects like AAVE, involve affected communities in development, and conduct regular bias audits. Systems should offer user feedback options, support contextual understanding, and be built by culturally aware teams.