@inproceedings{huang-etal-2025-nptu,
    title = "The {NPTU} {ASR} System for {FSR}2025 {H}akka Character/{P}inyin Recognition: Whisper with m{BART} Post-Editing and {RNNLM} Rescoring",
    author = "Huang, Yi-Chin  and
      Chen, Yu-Heng  and
      Wang, Jian-Hua  and
      Wu, Hsiu-Chi  and
      Kuo, Chih-Chung  and
      Huang, Chao-Shih  and
      Liao, Yuan-Fu",
    editor = "Chang, Kai-Wei  and
      Lu, Ke-Han  and
      Yang, Chih-Kai  and
      Tam, Zhi-Rui  and
      Chang, Wen-Yu  and
      Wang, Chung-Che",
    booktitle = "Proceedings of the 37th Conference on Computational Linguistics and Speech Processing (ROCLING 2025)",
    month = nov,
    year = "2025",
    address = "National Taiwan University, Taipei City, Taiwan",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/dashboard/2025.rocling-main.63/",
    pages = "518--522",
    ISBN = "979-8-89176-379-1",
    abstract = "This paper presents our system for the FSR-2025 Hakka Automatic Speech Recognition (ASR) Challenge, which consists of two sub-tasks: (i) Hakka Characters and (ii) Hakka Pinyin. We propose a unified architecture built upon Whisper [1], a large weakly supervised ASR model, as the acoustic backbone, with optional LoRA (Low-Rank Adaptation [2]) for parameter-efficient fine-tuning. Data augmentation techniques include the MUSAN [3] corpus (music/speech/noise) and tempo/speed perturbation [4]. For the character task, mBART-50 [5,6], a multilingual sequence-to-sequence model, is applied for text correction, while both tasks employ an RNNLM [7] for N-best rescoring. Under the final evaluation setting of the character task, mBART-driven 10-best text correction combined with RNNLM rescoring achieved a CER (Character Error Rate) of 6.26{\%}, whereas the official leaderboard reported 22.5{\%}. For the Pinyin task, the Medium model proved more suitable than the Large model given the dataset size and accent distribution. With 10-best RNNLM rescoring, it achieved a SER (Syllable Error Rate) of 4.65{\%} on our internal warm-up test set, and the official final score (with tone information) was 14.81{\%}. Additionally, we analyze the contribution of LID (Language Identification) for accent recognition across different recording and media sources."
}