Scholarlink

<p>In the process of food quality inspection, the real-time collection and precise record are essential for product quality control and traceability. However, traditional manual record methods are inefficient and error-prone. Currently, multi-modal perception has emerged as a key trend in data acquisition, in which speech information has shown considerable potential in complex field environments due to its natural, efficient, and contact-free characteristics. To address the limited adaptability of strong background noise, diverse accents, and dense domain-specific terminology in current general-purpose ASR models in quality inspection scenarios, this study focuses on aquatic product quality inspection and has constructed the HYAF speech corpus, consisting of 21,721 samples, which can provide a realistic and domain-relevant dataset for method validation. The FRCRN model has also been employed to enhance the SNR of speech inputs, improving average SNR from –2.76dB to 5.06dB. Then, we proposed a domain adaptation approach for Whisper based on SDFT, which generates soft labels through structured distillation templates to effectively mitigate distributional bias and catastrophic forgetting. Experimental results have shown that the proposed method can reduce the WER on the test set to 10.8%, shorten training time by 42.7% compared with conventional fine-tuning, and demonstrate superior domain adaptation and generalization capabilities under low-resource and high-noise conditions.</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

Fine-tuning Whisper for speech recognition in aquatic product inspection tasks

  • 期刊论文

  • 发表日期：

    2026年3月1日

查看全文
  • Ming Chen,
  • Zhanwang Gu,
  • Wenjuan Wang,
  • Lina Pei,
  • Yan Ge,
  • Yibo Zou

摘要

In the process of food quality inspection, the real-time collection and precise record are essential for product quality control and traceability. However, traditional manual record methods are inefficient and error-prone. Currently, multi-modal perception has emerged as a key trend in data acquisition, in which speech information has shown considerable potential in complex field environments due to its natural, efficient, and contact-free characteristics. To address the limited adaptability of strong background noise, diverse accents, and dense domain-specific terminology in current general-purpose ASR models in quality inspection scenarios, this study focuses on aquatic product quality inspection and has constructed the HYAF speech corpus, consisting of 21,721 samples, which can provide a realistic and domain-relevant dataset for method validation. The FRCRN model has also been employed to enhance the SNR of speech inputs, improving average SNR from –2.76dB to 5.06dB. Then, we proposed a domain adaptation approach for Whisper based on SDFT, which generates soft labels through structured distillation templates to effectively mitigate distributional bias and catastrophic forgetting. Experimental results have shown that the proposed method can reduce the WER on the test set to 10.8%, shorten training time by 42.7% compared with conventional fine-tuning, and demonstrate superior domain adaptation and generalization capabilities under low-resource and high-noise conditions.