Model-centric data selection: Refining end-to-end speech recognition

Abstract

Data selection can be an important step in pre-processing datasets for Automatic Speech Recognition (ASR) -- still its application is not general. In order to handle potential labeling errors and other anomalies in the dataset, we introduced a simple model-centric speech data selection strategy. It discards samples in the dataset that is difficult to recognize by the model, and use a restricted dataset to retrain the model. This technique improved the recognition accuracy of Hungarian ASR both on the BEA-Base and Common Voice (CV) datasets by using the Conformer model architecture. The proposed approach achieved a consistent relative improvement in terms of both Character and Word Error Rates (CER, WER), up to (3%, 2.5%).