@inproceedings{haider-etal-2026-improving,
title = "Improving {U}rdu {ASR} with Fine-Tuned Small Language Models for Post-{ASR} Error Correction",
author = "Haider, Sajjad and
Sheikh, Solat J. and
Aqib, Zuha and
Inayat, Farah and
Ahmed, Zehra",
editor = "T.Y.S.S., Santosh and
Ghosh, Sreyan and
Duraiswami, Ramani and
Watanabe, Shinji and
Arora, Siddhant and
Kumar, Sonal and
Adi, Yossi and
Manocha, Dinesh and
Garcia Perera, Leibny Paola",
booktitle = "The Proceedings of Speech and Audio Language Models Workshop (2nd Edition)",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.salma-1.7/",
pages = "74--84",
ISBN = "979-8-89176-496-5",
abstract = "Current multilingual automatic speech recognition (ASR) models have substantially improved speech recognition across many languages, yet recognition quality remains limited for low-resource languages such as Urdu. A common approach to improve recognition quality is to fine-tune the ASR model itself, but this requires labelled speech data and retraining the recognizer. A recently emerging alternative is Generative Error Correction (GEC), in which a language model post-processes ASR output to correct transcription errors while leaving the underlying speech recognizer unchanged. This paper investigates this approach for Urdu by keeping Whisper-Turbo frozen and fine-tuning small language models to correct its transcription errors. To address data scarcity in Urdu, we construct a pseudo-gold training dataset using a strong teacher model. Fine-tuning 3.5{--}4B parameter language models on this data achieves a {\textasciitilde}33{\%} relative WER reduction (from 11.74{\%} to 7.81{\%}), outperforming zero-shot baselines as well as significantly larger 12{--}14B zero-shot models. These results show that combining a frozen recognizer with a small fine-tuned corrector provides a practical, resource-efficient framework for Urdu ASR improvement without relying on ASR retraining or substantially larger models."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="haider-etal-2026-improving">
<titleInfo>
<title>Improving Urdu ASR with Fine-Tuned Small Language Models for Post-ASR Error Correction</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sajjad</namePart>
<namePart type="family">Haider</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Solat</namePart>
<namePart type="given">J</namePart>
<namePart type="family">Sheikh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zuha</namePart>
<namePart type="family">Aqib</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Farah</namePart>
<namePart type="family">Inayat</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zehra</namePart>
<namePart type="family">Ahmed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>The Proceedings of Speech and Audio Language Models Workshop (2nd Edition)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Santosh</namePart>
<namePart type="family">T.Y.S.S.</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sreyan</namePart>
<namePart type="family">Ghosh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ramani</namePart>
<namePart type="family">Duraiswami</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shinji</namePart>
<namePart type="family">Watanabe</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siddhant</namePart>
<namePart type="family">Arora</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sonal</namePart>
<namePart type="family">Kumar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yossi</namePart>
<namePart type="family">Adi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dinesh</namePart>
<namePart type="family">Manocha</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Leibny</namePart>
<namePart type="given">Paola</namePart>
<namePart type="family">Garcia Perera</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-496-5</identifier>
</relatedItem>
<abstract>Current multilingual automatic speech recognition (ASR) models have substantially improved speech recognition across many languages, yet recognition quality remains limited for low-resource languages such as Urdu. A common approach to improve recognition quality is to fine-tune the ASR model itself, but this requires labelled speech data and retraining the recognizer. A recently emerging alternative is Generative Error Correction (GEC), in which a language model post-processes ASR output to correct transcription errors while leaving the underlying speech recognizer unchanged. This paper investigates this approach for Urdu by keeping Whisper-Turbo frozen and fine-tuning small language models to correct its transcription errors. To address data scarcity in Urdu, we construct a pseudo-gold training dataset using a strong teacher model. Fine-tuning 3.5–4B parameter language models on this data achieves a ~33% relative WER reduction (from 11.74% to 7.81%), outperforming zero-shot baselines as well as significantly larger 12–14B zero-shot models. These results show that combining a frozen recognizer with a small fine-tuned corrector provides a practical, resource-efficient framework for Urdu ASR improvement without relying on ASR retraining or substantially larger models.</abstract>
<identifier type="citekey">haider-etal-2026-improving</identifier>
<location>
<url>https://aclanthology.org/2026.salma-1.7/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>74</start>
<end>84</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Improving Urdu ASR with Fine-Tuned Small Language Models for Post-ASR Error Correction
%A Haider, Sajjad
%A Sheikh, Solat J.
%A Aqib, Zuha
%A Inayat, Farah
%A Ahmed, Zehra
%Y T.Y.S.S., Santosh
%Y Ghosh, Sreyan
%Y Duraiswami, Ramani
%Y Watanabe, Shinji
%Y Arora, Siddhant
%Y Kumar, Sonal
%Y Adi, Yossi
%Y Manocha, Dinesh
%Y Garcia Perera, Leibny Paola
%S The Proceedings of Speech and Audio Language Models Workshop (2nd Edition)
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-496-5
%F haider-etal-2026-improving
%X Current multilingual automatic speech recognition (ASR) models have substantially improved speech recognition across many languages, yet recognition quality remains limited for low-resource languages such as Urdu. A common approach to improve recognition quality is to fine-tune the ASR model itself, but this requires labelled speech data and retraining the recognizer. A recently emerging alternative is Generative Error Correction (GEC), in which a language model post-processes ASR output to correct transcription errors while leaving the underlying speech recognizer unchanged. This paper investigates this approach for Urdu by keeping Whisper-Turbo frozen and fine-tuning small language models to correct its transcription errors. To address data scarcity in Urdu, we construct a pseudo-gold training dataset using a strong teacher model. Fine-tuning 3.5–4B parameter language models on this data achieves a ~33% relative WER reduction (from 11.74% to 7.81%), outperforming zero-shot baselines as well as significantly larger 12–14B zero-shot models. These results show that combining a frozen recognizer with a small fine-tuned corrector provides a practical, resource-efficient framework for Urdu ASR improvement without relying on ASR retraining or substantially larger models.
%U https://aclanthology.org/2026.salma-1.7/
%P 74-84
Markdown (Informal)
[Improving Urdu ASR with Fine-Tuned Small Language Models for Post-ASR Error Correction](https://aclanthology.org/2026.salma-1.7/) (Haider et al., SALMA 2026)
ACL