Race and Regionality on the ASpIRE ASR Model

Automatic speech recognition (ASR) enables the transcription of spoken speech into a written format. Previous research has shown racial biases in modern ASR systems exist and negatively affect Black speakers. In this thesis, speech data from the CallHome and CORAAL ATL, DCB, PRV, and ROC corpora are processed and given to ASpIRE, a DNN-HMM model built on the open-source ASR toolkit Kaldi. The trends in the model’s word error rates between different phonological phenomena and corpora are considered in the context of the model’s original training process and modern sociolinguistic knowledge. All in all, the training set used to develop the ASpIRE model is insufficiently enriched with phonological and lexical representations of AAL and Southern characteristics.