Pruned CTC for Memory-Efficient Large-Vocabulary ASR Training

Researchers present pruned CTC, a method that avoids materializing frame-by-vocabulary activations in memory to make large-vocabulary ASR training feasible. Conventional CTC implementations materialize these activations, which the work states makes CTC training with native LLM vocabularies prohibitively memory-intensive. The approach targets memory-efficient training of connectionist temporal classification models that support both offline and streaming speech recognition under utterance-level supervision.

Connectionist temporal classification CTC naturally supports offline and streaming speech recognition with utterance-level supervision, but conventional implementations materialize frame-by-vocabulary activations in memory, making CTC training with native LLM vocabularies prohibitively memory-inte