Clustering and mining accented speech for inclusive and fair speech recognition

A system comprising:

• data processing hardware;
• and
• memory hardware in communication with the data processing hardware and storing instructions that cause the data processing hardware to perform operations comprising:
• receiving a corpus of training utterances spoken across various accents, each training utterance in the corpus comprising training audio features characterizing the training utterance;
• and
• executing a training process to train an accent recognition model on the corpus of training utterances to teach the accent recognition model to learn how to predict accent representations from the training audio features, the accent recognition model comprising:
• one or more strided convolutional layers configured to sub-sample the training audio features of each training utterance and output a strided convolution output;
• a stack of multi-head attention layers configured to receive each strided convolution output from the strided convolutional layers and generate a final attention layer output at a last multi-head attention layer in the stack of multi-head attention layers;
• and
• a pooling layer configured to receive the final attention layer output and pool an output for each training utterance across time to generate a corresponding accent representation.