QuaSR: New framework improves speech recognition for Pacific Indigenous languages

Training ASR models for low-resource languages like those in the Pacific Islands is notoriously difficult due to limited data and inconsistent supervision quality. Noisy audio, transcription errors, and unreliable alignments often derail fine-tuning. To tackle this, researchers from the University of Melbourne and Australian National University introduce QuaSR (Quality-Aware Sample Reweighting), a framework that scores each training sample by combining two complementary signals: data-side reliability and model-side learnability.

Data reliability is estimated from three dimensions: acoustic quality, transcription consistency, and alignment confidence. Meanwhile, learnability is captured via the model's training loss—samples the model finds hard to learn are weighted differently. These scores are integrated into a unified utility measure that guides sample weighting during adaptation. The team tested QuaSR on four Pacific Indigenous languages and found that the utility scores strongly correlate with downstream ASR performance.

Compared to standard fine-tuning and other data selection methods, QuaSR consistently improved word error rates across all languages. The approach is simple yet effective, requiring no architectural changes—just a weighting mechanism during training. This opens a new path for leveraging difficulty scores in low-resource speech learning, especially for endangered languages where every data point matters.