Scholarlink

<p>In saturation diving, reliable speech communication under helium–oxygen (Heliox) conditions is critical for operational safety and efficiency. Heliox speech exhibits severe acoustic mismatch relative to standard air speech, and recognition performance further degrades in the presence of chamber/environmental noise and domain-specific terminology. To study this problem in a realistic setting, we collected Heliox speech recordings at two saturation conditions (12 m and 25 m equivalent depths) and constructed a corresponding dataset. We then adapt Whisper-large-v3 via Low-Rank Adaptation (LoRA) to enable parameter-efficient domain adaptation, and enhance decoding using practical inference-time components, including hotword biasing, language-model (LM) reranking, test-time augmentation (TTA) with speed perturbation, and rolling context prompts, together with chunked decoding for stable deployment. On our Heliox evaluation sets, the proposed system achieved a character error rate (CER) of 4.725% at a water depth of 12 m and a CER of 7.165% at a water depth of 25 m, under the reported decoding configuration, while maintaining practical inference cost on GPU/CPU server platforms. We note that inference-time strategies provide complementary robustness gains but do not fully eliminate the need for domain adaptation under severe Heliox-induced shifts.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LoRA-enhanced whisper for resource-efficient heliox speech recognition

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月18日

查看全文
  • Weichang Mao,
  • Haojie Gu,
  • Jia He,
  • Yu Li,
  • Shifeng Wang

摘要

In saturation diving, reliable speech communication under helium–oxygen (Heliox) conditions is critical for operational safety and efficiency. Heliox speech exhibits severe acoustic mismatch relative to standard air speech, and recognition performance further degrades in the presence of chamber/environmental noise and domain-specific terminology. To study this problem in a realistic setting, we collected Heliox speech recordings at two saturation conditions (12 m and 25 m equivalent depths) and constructed a corresponding dataset. We then adapt Whisper-large-v3 via Low-Rank Adaptation (LoRA) to enable parameter-efficient domain adaptation, and enhance decoding using practical inference-time components, including hotword biasing, language-model (LM) reranking, test-time augmentation (TTA) with speed perturbation, and rolling context prompts, together with chunked decoding for stable deployment. On our Heliox evaluation sets, the proposed system achieved a character error rate (CER) of 4.725% at a water depth of 12 m and a CER of 7.165% at a water depth of 25 m, under the reported decoding configuration, while maintaining practical inference cost on GPU/CPU server platforms. We note that inference-time strategies provide complementary robustness gains but do not fully eliminate the need for domain adaptation under severe Heliox-induced shifts.