Targeted Subset Selection for Limited-data ASR Accent Adaptation

We study the task of adapting an existing ASR model to a non-native accent while being constrained by a transcription budget on the duration of utterances selected from a large unlabeled corpus. We propose a subset selection approach using the recently proposed submodular mutual information functions, in which we identify a diverse set of utterances that match the target accent. This is specified through a few target utterances and achieved by modelling the relationship between the target and the selected subsets using these functions. The model adapts to the accent through fine-tuning with utterances selected and transcribed from the unlabeled corpus. We also use an accent classifier to learn accent-aware feature representations. Our method is also able to exploit samples from other accents to perform out-of-domain selections for low-resource accents which are not available in these corpora. We show that the targeted subset selection approach improves significantly upon random sampling - by around 5% to 10% (absolute) in most cases, and is around 10x more label-efficient. We also compare with an oracle method where we specifically pick from the target accent and our method is comparable to the oracle in its selections and WER performance.