The ASR Post-Processor Performance Challenges of BackTranScription (BTS): Data-Centric and Model-Centric Approaches

Training an automatic speech recognition (ASR) post-processor based on sequence-to-sequence (S2S) requires a parallel pair (e.g., result and human post-edited sentence) to construct the dataset, which demands great amount of labor. BackTransScription (BTS) proposes data-building method mitigate limitations existing S2S ASR post-processors, can automatically generate vast amounts training datasets, reducing time cost in data construction. Despite emergence this novel approach, BTS-based still has research challenges is mostly untested diverse approaches. In study, we highlight these through detailed experiments by analyzing data-centric approach (i.e., controlling without model alteration) model-centric modification). other words, attempt point out problems with current trend pursuing alert against ignoring importance data. Our experiment results show that outperformed +11.69, +17.64, +19.02 F1-score, BLEU, GLEU tests.