ASR filtering method cuts error rates by 6.4% on 90k-hour dataset

Researchers Kohei Matsuura and Masato Mimura have introduced a novel training approach for automatic speech recognition (ASR) that better leverages large-scale weakly supervised datasets. Such datasets are essential for training robust end-to-end models but often suffer from noisy labels and lack domain specificity. Their method addresses this through three sequential steps: initial pretraining on the entire 90,000-hour Japanese dataset, continued pretraining on a filtered subset selected by character error rate (CER), and finally fine-tuning on a small number of acoustically similar samples to the target domain, drawn from the filtered subset.

The results are striking: the filtering step alone reduced CER by 6.4%, and the selection step added another 4.0% improvement — a combined gain of up to 10.4% despite reusing samples already seen in the first pretraining. This demonstrates that careful data curation, even without new data, can significantly boost ASR performance. The paper, submitted to arXiv on June 27, 2026, provides a practical path for practitioners working with noisy, unsupervised audio collections to build more accurate speech systems without requiring manual transcription. The approach is especially valuable for low-resource languages or specialized domains where clean labeled data is scarce, showing that smarter training sequences can unlock hidden value in existing datasets.