Hakka Speech Recognition with Whisper and Pinyin Post-processing for FSR-2025

Chia-Hsin Lee, Yung-Jun Chang, Jin-Yan Wu, Kuan-Yu Chen


Abstract
本研究為參加 FSR-2025 客語語音辨識挑戰賽(Hakka ASR II)的技術報告,旨在推進客語自動語音辨識技術的發展。由於客語屬於低資源語言,且存在多種腔調,語音辨識面臨高度挑戰。我們以 Whisperlarge-v2 為骨幹模型,設計兩階段訓練流程:首先利用「Hakka Across Taiwan(HAT)」語料庫進行模型調適,以捕捉客語的一般聲學特徵;其次在賽事方提供的60 小時腔調語料上進行微調,以增強對目標資料的適應性。實驗發現,直接輸出客語漢字可達到良好的字錯率(CER),但由 於腔調差異與拼音規則變化多,拼音任務表現顯著下降。為解決此問題,我們以漢字模型的編碼器初始化拼音模型,並提出結合 RoBERTa 漢字轉拼音、腔調判斷與字典修正的後處理模組,期望可以在比賽中提升辨識的成效。
Anthology ID:
2025.rocling-main.53
Volume:
Proceedings of the 37th Conference on Computational Linguistics and Speech Processing (ROCLING 2025)
Month:
November
Year:
2025
Address:
National Taiwan University, Taipei City, Taiwan
Editors:
Kai-Wei Chang, Ke-Han Lu, Chih-Kai Yang, Zhi-Rui Tam, Wen-Yu Chang, Chung-Che Wang
Venue:
ROCLING
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
454–458
Language:
URL:
https://aclanthology.org/2025.rocling-main.53/
DOI:
Bibkey:
Cite (ACL):
Chia-Hsin Lee, Yung-Jun Chang, Jin-Yan Wu, and Kuan-Yu Chen. 2025. Hakka Speech Recognition with Whisper and Pinyin Post-processing for FSR-2025. In Proceedings of the 37th Conference on Computational Linguistics and Speech Processing (ROCLING 2025), pages 454–458, National Taiwan University, Taipei City, Taiwan. Association for Computational Linguistics.
Cite (Informal):
Hakka Speech Recognition with Whisper and Pinyin Post-processing for FSR-2025 (Lee et al., ROCLING 2025)
Copy Citation:
PDF:
https://aclanthology.org/2025.rocling-main.53.pdf