Standard CTC training stores frame-by-vocabulary activations, so using native LLM vocabularies becomes very memory intensive. Pruned CTC observes that every valid alignment uses only blank and the transcript tokens, so it runs the expensive alignment steps over that small subset while still normalizing over the full vocabulary. The authors prove this is exactly equivalent to full-vocabulary CTC in loss and first-order gradients, then add finite-beam alignment pruning as a further approximation.
With a Zipformer-M encoder and a 180K vocabulary, Pruned CTC cuts full-step memory by 5.1x with only 17% step-time overhead and matches standard CTC accuracy on three corpora. The same approach enables LLM-CTC, which adapts pretrained causal LLMs to non-autoregressive ASR using their native vocabularies. On GigaSpeech, Qwen3 models from 0.6B to 32B stay within 7% relative WER of LLM-CE while recognizing 7 to 10 times faster, and a bounded-history streaming variant stays within 3% relative WER of matched offline models.