Test-Time Vocabulary Expansion for Automatic Speech Recognition
September 25, 2026
This method enables ASR models to learn new, out-of-vocabulary words from unlabeled test data at test time. By combining a frozen CTC acoustic model with a language model and an adaptation module, the system optimizes lexical token representations using a Kullback-Leibler divergence objective.
HOW THIS AFFECTS YOU
●
builderThis technique can improve the robustness of your speech recognition pipelines against emerging terminology.
●
researcherYou can integrate test-time adaptation to allow models to handle evolving vocabularies without retraining.