@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://aclanthology.org/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."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="huang-etal-2025-nptu">
<titleInfo>
<title>The NPTU ASR System for FSR2025 Hakka Character/Pinyin Recognition: Whisper with mBART Post-Editing and RNNLM Rescoring</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yi-Chin</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yu-Heng</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jian-Hua</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hsiu-Chi</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chih-Chung</namePart>
<namePart type="family">Kuo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chao-Shih</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuan-Fu</namePart>
<namePart type="family">Liao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 37th Conference on Computational Linguistics and Speech Processing (ROCLING 2025)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kai-Wei</namePart>
<namePart type="family">Chang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ke-Han</namePart>
<namePart type="family">Lu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chih-Kai</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhi-Rui</namePart>
<namePart type="family">Tam</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wen-Yu</namePart>
<namePart type="family">Chang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chung-Che</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">National Taiwan University, Taipei City, Taiwan</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-379-1</identifier>
</relatedItem>
<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.</abstract>
<identifier type="citekey">huang-etal-2025-nptu</identifier>
<location>
<url>https://aclanthology.org/2025.rocling-main.63/</url>
</location>
<part>
<date>2025-11</date>
<extent unit="page">
<start>518</start>
<end>522</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T The NPTU ASR System for FSR2025 Hakka Character/Pinyin Recognition: Whisper with mBART Post-Editing and RNNLM Rescoring
%A Huang, Yi-Chin
%A Chen, Yu-Heng
%A Wang, Jian-Hua
%A Wu, Hsiu-Chi
%A Kuo, Chih-Chung
%A Huang, Chao-Shih
%A Liao, Yuan-Fu
%Y Chang, Kai-Wei
%Y Lu, Ke-Han
%Y Yang, Chih-Kai
%Y Tam, Zhi-Rui
%Y Chang, Wen-Yu
%Y Wang, Chung-Che
%S Proceedings of the 37th Conference on Computational Linguistics and Speech Processing (ROCLING 2025)
%D 2025
%8 November
%I Association for Computational Linguistics
%C National Taiwan University, Taipei City, Taiwan
%@ 979-8-89176-379-1
%F huang-etal-2025-nptu
%X 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.
%U https://aclanthology.org/2025.rocling-main.63/
%P 518-522
Markdown (Informal)
[The NPTU ASR System for FSR2025 Hakka Character/Pinyin Recognition: Whisper with mBART Post-Editing and RNNLM Rescoring](https://aclanthology.org/2025.rocling-main.63/) (Huang et al., ROCLING 2025)
ACL
- Yi-Chin Huang, Yu-Heng Chen, Jian-Hua Wang, Hsiu-Chi Wu, Chih-Chung Kuo, Chao-Shih Huang, and Yuan-Fu Liao. 2025. The NPTU ASR System for FSR2025 Hakka Character/Pinyin Recognition: Whisper with mBART Post-Editing and RNNLM Rescoring. In Proceedings of the 37th Conference on Computational Linguistics and Speech Processing (ROCLING 2025), pages 518–522, National Taiwan University, Taipei City, Taiwan. Association for Computational Linguistics.