AI
Pruned CTC for Memory-Efficient Large-Vocabulary ASR Training
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 26/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 03:27:38 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 15:18:24 UTC.
Briefings
This study introduces Pruned CTC, a method that reduces memory usage in Large Vocabulary ASR training by restricting alignment computation to a small subset of vocabulary tokens. The research proves this reduction is equivalent to full-vocabulary CTC in terms of loss and gradients. Experiments show Pruned CTC reduces full-step memory by 5.1 times while maintaining only 17% step-time overhead. The method was tested across three corpora and on GigaSpeech with six Qwen3 model sizes, demonstrating that LLM-CTC achieves within 7% relative WER of LLM-CE with significantly faster recognition speeds.
Key points
- Pruned CTC restricts alignment computation to a small subset of vocabulary tokens, reducing memory usage by 5.1 times compared to conventional implementations.
- The method maintains full-vocabulary normalization and proves equivalence to full-vocabulary CTC in loss and gradients, enabling efficient training with native LLM vocabularies.
Context and limitations
Results are limited to offline and streaming settings; specific performance metrics for all six Qwen3 model sizes on GigaSpeech are not fully detailed, and the 17% step-time overhead represents a trade-off for the 5.1x memory reduction.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 03:27:38 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifz3or2ag4vp76hbqxkegefw7dr77xjvqxymz6bg2rgirabkoo6oe