Analysis of CTC and neural transducer models shows that fixed-size beam search causes hypothesis count spikes during periods of model uncertainty. Implementing score-based pruning dynamically adapts the search space, significantly improving Real-Time Factor (RTF) efficiency.
HOW THIS AFFECTS YOU
●
builderYou can improve ASR inference latency by replacing fixed beam search with uncertainty-aware pruning.
●
researcherThis provides a framework for analyzing search space complexity in end-to-end sequence-to-sequence architectures.