Sample-Efficient ASR Training via Utterance Selection

The patent extracts and identifies a small subset of high-value training utterances from the overall training data pool. By using importance sampling, the system extracts only the most informative samples that contribute most to reducing word-error rate, rather than using all available training data. This extraction principle directly addresses the contradiction by achieving good recognition accuracy with a much smaller, carefully selected subset of training samples.

The patent applies local quality by treating different training utterances differently based on their individual value. Instead of uniformly processing all training data, the system assigns different importance weights to different utterances based on their contribution to model improvement. High-value utterances (those with higher importance scores) are selected for annotation and training, while low-value utterances are discarded. This selective approach resolves the contradiction by focusing resources on the most impactful samples.

The patent extracts only the most valuable training utterances that need manual annotation. By calculating importance scores for each utterance based on the ASR system's current performance and uncertainty, the system identifies a small subset of high-impact samples. Annotation efforts are concentrated only on these extracted high-value samples rather than annotating all training data, thereby significantly reducing annotation time while maintaining or improving recognition quality.

The patent implements a self-service mechanism where the ASR system automatically identifies which utterances would benefit most from annotation. The system uses its own performance metrics (word-error rate, confidence scores) to self-determine the priority of different training samples. This eliminates the need for manual assessment of training data value and automates the selection process, reducing human labor time while improving annotation efficiency.

When new utterances need to be recognized, the patent extracts only the most informative new training samples rather than annotating all new data. The importance sampling mechanism identifies which new utterances provide the most value for expanding the system's recognition capability. This selective extraction allows the system to adapt to new speech patterns and vocabulary with minimal annotation resources.

The patent applies local quality by differentiating between various types of new training utterances based on their specific contribution to adaptability. Not all new utterances are equally valuable for improving recognition capability. The system identifies and prioritizes annotation of specific high-value new samples (those with high importance scores) while deprioritizing or skipping low-value new samples, thereby achieving adaptability improvements with reduced annotation resources.