NOT hearing African American Vernacular English (AAVE)
NOT hearing African American Vernacular English (AAVE):
As shown by the trial of George Zimmerman for the murder of Trayvon Martin and the errors in the Automated Speech Recognition (ASR) systems used by Amazon, Apple, Microsoft, Google and IBM
John R. Rickford
J.E. Wallace Sterling Professor of Linguistics and the Humanities
March 24, 2022
5:00 - 6:00 pm
139 Social Sciences and by Zoom
This is a hybrid event with limited in-person capacity. Registration is required for attending on Zoom and in-person.
in person: https://duke.qualtrics.com/jfe/form/SV_86vAoiWpqSeEW9M
African American Vernacular English (AAVE) is by far the most studied variety of American English, yet most non-linguists either ignore or deny it, or notice it only when it is the source of national controversies in education, as it was in the 1996 Oakland Ebonics firestorm.
Recently, however, it became an issue of concern when Koenecke et al (2020) showed that the Automated Speech Recognition (ASR) systems used by Amazon, Apple, Google, IBM and Microsoft had on average twice as many errors when transcribing the speech of black speakers as they did when transcribing the speech of white speakers.
In this talk Professor Rickford will discuss what AAVE is and explain some of the challenges its pronunciation and grammar represent for the ASRs of Apple and other systems. He'll actually begin with a parallel example of the difficulties that AAVE as spoken by vernacular black speakers pose for jurors and court reporters, drawing on a sample of the speech of Rachel Jeantel of Miami (Rickford and King 2016). Rachel was a good friend of Trayvon Martin and was the prosecution star witness in the trial of George Zimmerman for Trayvon's murder in 2012. Discussing her speech will demonstrate some of AAVE's features and demonstrate that they are systematic and rule-governed, as is true of ALL languages and dialects.
From there we will proceed to examples of actual speech snippets used to test the ASRs of Apple and other systems, illustrating some of the specific difficulties they caused. We will also introduce limited evidence of difficulties that ASRs faced with Latino English and discuss possible solutions to these recurrent limitations with ASR devices.