Multimodal Consistency-Guided Reference-Free Data Selection for ASR Accent Adaptation

Researchers have developed a new AI pipeline that dramatically reduces the data needed to train speech recognition models for different accents. Their method selects the most useful unlabeled audio data by checking consistency between speech and generated text. Using just 1,500 carefully chosen utterances from a pool of 30,000, they achieved a 10.91% word error rate—nearly matching the 10.45% rate achieved using all 30,000 fully labeled examples.