@inproceedings{pandit-etal-2026-reward,
title = "Reward-Guided Fine-Tuning of Whisper for Low-Resource {N}epali Speech Recognition",
author = "Pandit, Aadarsh and
Khanal, Yudhin and
Pandey, Ishan and
Kunwar, Kushal and
Regmi, Sunil",
editor = "Sarveswaran, Kengatharaiyer and
Vaidya, Ashwini",
booktitle = "Proceedings of the Second workshop on Challenges in Processing {S}outh {A}sian Languages ({CH}i{PSAL}2026)",
month = may,
year = "2026",
address = "Palma de Mallorca, Spain",
publisher = "ELRA Language Resources Association (ELRA)",
url = "https://aclanthology.org/2026.chipsal-1.16/",
doi = "10.63317/45u9hjmchyhh",
pages = "169--175",
abstract = "Fine tuning speech recognition models on noisy real world data is tricky. The model has no way of knowing which training samples are reliable and which are not, so it ends up learning from bad examples just as readily as good ones. This is a real problem for Nepali, where most available training data comes from YouTube videos with automatically generated subtitles that are often inaccurate. In this work, we tried a simple fix. Instead of feeding everything to the model, we first asked humans to rate the quality of a sample of transcriptions, trained a small Random Forest classifier on those 2,000 ratings, and used it to filter out the bad samples before each retraining round. The classifier uses four automatically computable features, Word Error Rate (WER), Character Error Rate (CER), length ratio, and length difference, and achieves 81{\%} accuracy on a held out set. Running two filtering and retraining cycles on a 40,000 clip training subset drawn from a 68.4 hour corpus improves substantially over our own standard fine tuning baseline of 5.60{\%} WER and 5.10{\%} CER, reaching 4.89{\%} WER and 4.52{\%} CER, which corresponds to an 11 to 13{\%} relative gain. The approach is much lighter than full Reinforcement Learning from Human Feedback but still uses real human judgment to guide training."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="pandit-etal-2026-reward">
<titleInfo>
<title>Reward-Guided Fine-Tuning of Whisper for Low-Resource Nepali Speech Recognition</title>
</titleInfo>
<name type="personal">
<namePart type="given">Aadarsh</namePart>
<namePart type="family">Pandit</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yudhin</namePart>
<namePart type="family">Khanal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ishan</namePart>
<namePart type="family">Pandey</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kushal</namePart>
<namePart type="family">Kunwar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunil</namePart>
<namePart type="family">Regmi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Second workshop on Challenges in Processing South Asian Languages (CHiPSAL2026)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kengatharaiyer</namePart>
<namePart type="family">Sarveswaran</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ashwini</namePart>
<namePart type="family">Vaidya</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA Language Resources Association (ELRA)</publisher>
<place>
<placeTerm type="text">Palma de Mallorca, Spain</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Fine tuning speech recognition models on noisy real world data is tricky. The model has no way of knowing which training samples are reliable and which are not, so it ends up learning from bad examples just as readily as good ones. This is a real problem for Nepali, where most available training data comes from YouTube videos with automatically generated subtitles that are often inaccurate. In this work, we tried a simple fix. Instead of feeding everything to the model, we first asked humans to rate the quality of a sample of transcriptions, trained a small Random Forest classifier on those 2,000 ratings, and used it to filter out the bad samples before each retraining round. The classifier uses four automatically computable features, Word Error Rate (WER), Character Error Rate (CER), length ratio, and length difference, and achieves 81% accuracy on a held out set. Running two filtering and retraining cycles on a 40,000 clip training subset drawn from a 68.4 hour corpus improves substantially over our own standard fine tuning baseline of 5.60% WER and 5.10% CER, reaching 4.89% WER and 4.52% CER, which corresponds to an 11 to 13% relative gain. The approach is much lighter than full Reinforcement Learning from Human Feedback but still uses real human judgment to guide training.</abstract>
<identifier type="citekey">pandit-etal-2026-reward</identifier>
<identifier type="doi">10.63317/45u9hjmchyhh</identifier>
<location>
<url>https://aclanthology.org/2026.chipsal-1.16/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>169</start>
<end>175</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Reward-Guided Fine-Tuning of Whisper for Low-Resource Nepali Speech Recognition
%A Pandit, Aadarsh
%A Khanal, Yudhin
%A Pandey, Ishan
%A Kunwar, Kushal
%A Regmi, Sunil
%Y Sarveswaran, Kengatharaiyer
%Y Vaidya, Ashwini
%S Proceedings of the Second workshop on Challenges in Processing South Asian Languages (CHiPSAL2026)
%D 2026
%8 May
%I ELRA Language Resources Association (ELRA)
%C Palma de Mallorca, Spain
%F pandit-etal-2026-reward
%X Fine tuning speech recognition models on noisy real world data is tricky. The model has no way of knowing which training samples are reliable and which are not, so it ends up learning from bad examples just as readily as good ones. This is a real problem for Nepali, where most available training data comes from YouTube videos with automatically generated subtitles that are often inaccurate. In this work, we tried a simple fix. Instead of feeding everything to the model, we first asked humans to rate the quality of a sample of transcriptions, trained a small Random Forest classifier on those 2,000 ratings, and used it to filter out the bad samples before each retraining round. The classifier uses four automatically computable features, Word Error Rate (WER), Character Error Rate (CER), length ratio, and length difference, and achieves 81% accuracy on a held out set. Running two filtering and retraining cycles on a 40,000 clip training subset drawn from a 68.4 hour corpus improves substantially over our own standard fine tuning baseline of 5.60% WER and 5.10% CER, reaching 4.89% WER and 4.52% CER, which corresponds to an 11 to 13% relative gain. The approach is much lighter than full Reinforcement Learning from Human Feedback but still uses real human judgment to guide training.
%R 10.63317/45u9hjmchyhh
%U https://aclanthology.org/2026.chipsal-1.16/
%U https://doi.org/10.63317/45u9hjmchyhh
%P 169-175
Markdown (Informal)
[Reward-Guided Fine-Tuning of Whisper for Low-Resource Nepali Speech Recognition](https://aclanthology.org/2026.chipsal-1.16/) (Pandit et al., CHiPSAL 2026)
ACL