Low-Resource Vietnamese Logistics Outbound-Call Speech Recognition with Low-Rank Adaptation and Dynamic Curriculum Learning

Reliable manual annotations for Vietnamese logistics calls are scarce, and telephone channel degradation compounds errors in business-critical entities. Existing adaptation methods typically optimize aggregate transcription accuracy, use uniform token losses, or apply a fixed training schedule; few control parameter updates, entity-level supervision, and sample order together when labels are noisy. We address this problem by combining low-rank adaptation (LoRA), token-level entity-weighted cross-entropy, and validation feedback curriculum learning for Whisper-large-v3. A joint acoustic–semantic score ranks samples by signal quality, sequence length, entity density, and linguistic rarity. The curriculum boundary expands only after the validation loss improves. On 20,317 in-domain training utterances, the complete system achieves a word error rate (WER) of 0.0735 and a mean exact match entity accuracy of 92.1%. Compared with vanilla LoRA, it reduces WER by 4.92% and raises entity accuracy from 85.3% to 92.1%. Ablation results show that curriculum pacing and entity weighting each contribute to the final performance.