@inproceedings{abu-horaira-chowdhury-2026-eraserhead,
title = "Eraserhead at {OSACT}7 Shared Task: {ASR} Consistency Filtering and Speaker-Adaptive Post-Processing for {A}rabic Speech Diacritization",
author = "Abu Horaira, Muhammad and
Chowdhury, Nahian",
editor = "Al-Khalifa, Hend and
El-Haj, Mo and
Ezzini, Saad",
booktitle = "The 7th Workshop on Open-Source {A}rabic Corpora and Processing Tools ({OSACT}7) with 5 Shared Tasks",
month = may,
year = "2026",
address = "Palma, Mallorca (Spain)",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.osact-1.32/",
doi = "10.63317/4x2nvon8vija",
pages = "247--251",
abstract = "Arabic speech diacritization is the task of restoring short vowel marks to undiacritized text derived from speech input. It remains difficult because ASR output can be noisy, dialectal variation is substantial, and speakers often differ in how they realize word-final diacritics. In this paper, we describe our submission to Task 2 of the KSAA-2026 Shared Task on Arabic Speech Dictation with Automatic Diacritization, where our system ranked 4th on the official leaderboard. Our approach builds on a pretrained ASR-aware diacritization model and adds three components: ASR Consistency Filtering, confidence-based ensembling of three checkpoints, and speaker-adaptive post-processing specifically for word-final diacritics. Rather than discarding problematic data, our filtering strategy replaces unreliable ASR transcripts with the undiacritized gold text rather than removing training examples, which makes training more stable. On the official test set, our system achieved a Diacritic Error Rate (DER) of 8.23, a Word Error Rate (WER) of 30.37, and a Sentence Error Rate (SER) of 80.79 under the With Case Endings (WCE), Including No Diacritic (Incl. 0) evaluation setting. It also outperformed the organizers' fine-tuned Text+ASR baseline in three of the four main evaluation settings."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="abu-horaira-chowdhury-2026-eraserhead">
<titleInfo>
<title>Eraserhead at OSACT7 Shared Task: ASR Consistency Filtering and Speaker-Adaptive Post-Processing for Arabic Speech Diacritization</title>
</titleInfo>
<name type="personal">
<namePart type="given">Muhammad</namePart>
<namePart type="family">Abu Horaira</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nahian</namePart>
<namePart type="family">Chowdhury</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>The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hend</namePart>
<namePart type="family">Al-Khalifa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mo</namePart>
<namePart type="family">El-Haj</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Saad</namePart>
<namePart type="family">Ezzini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Palma, Mallorca (Spain)</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Arabic speech diacritization is the task of restoring short vowel marks to undiacritized text derived from speech input. It remains difficult because ASR output can be noisy, dialectal variation is substantial, and speakers often differ in how they realize word-final diacritics. In this paper, we describe our submission to Task 2 of the KSAA-2026 Shared Task on Arabic Speech Dictation with Automatic Diacritization, where our system ranked 4th on the official leaderboard. Our approach builds on a pretrained ASR-aware diacritization model and adds three components: ASR Consistency Filtering, confidence-based ensembling of three checkpoints, and speaker-adaptive post-processing specifically for word-final diacritics. Rather than discarding problematic data, our filtering strategy replaces unreliable ASR transcripts with the undiacritized gold text rather than removing training examples, which makes training more stable. On the official test set, our system achieved a Diacritic Error Rate (DER) of 8.23, a Word Error Rate (WER) of 30.37, and a Sentence Error Rate (SER) of 80.79 under the With Case Endings (WCE), Including No Diacritic (Incl. 0) evaluation setting. It also outperformed the organizers’ fine-tuned Text+ASR baseline in three of the four main evaluation settings.</abstract>
<identifier type="citekey">abu-horaira-chowdhury-2026-eraserhead</identifier>
<identifier type="doi">10.63317/4x2nvon8vija</identifier>
<location>
<url>https://aclanthology.org/2026.osact-1.32/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>247</start>
<end>251</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Eraserhead at OSACT7 Shared Task: ASR Consistency Filtering and Speaker-Adaptive Post-Processing for Arabic Speech Diacritization
%A Abu Horaira, Muhammad
%A Chowdhury, Nahian
%Y Al-Khalifa, Hend
%Y El-Haj, Mo
%Y Ezzini, Saad
%S The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks
%D 2026
%8 May
%I Association for Computational Linguistics
%C Palma, Mallorca (Spain)
%F abu-horaira-chowdhury-2026-eraserhead
%X Arabic speech diacritization is the task of restoring short vowel marks to undiacritized text derived from speech input. It remains difficult because ASR output can be noisy, dialectal variation is substantial, and speakers often differ in how they realize word-final diacritics. In this paper, we describe our submission to Task 2 of the KSAA-2026 Shared Task on Arabic Speech Dictation with Automatic Diacritization, where our system ranked 4th on the official leaderboard. Our approach builds on a pretrained ASR-aware diacritization model and adds three components: ASR Consistency Filtering, confidence-based ensembling of three checkpoints, and speaker-adaptive post-processing specifically for word-final diacritics. Rather than discarding problematic data, our filtering strategy replaces unreliable ASR transcripts with the undiacritized gold text rather than removing training examples, which makes training more stable. On the official test set, our system achieved a Diacritic Error Rate (DER) of 8.23, a Word Error Rate (WER) of 30.37, and a Sentence Error Rate (SER) of 80.79 under the With Case Endings (WCE), Including No Diacritic (Incl. 0) evaluation setting. It also outperformed the organizers’ fine-tuned Text+ASR baseline in three of the four main evaluation settings.
%R 10.63317/4x2nvon8vija
%U https://aclanthology.org/2026.osact-1.32/
%U https://doi.org/10.63317/4x2nvon8vija
%P 247-251
Markdown (Informal)
[Eraserhead at OSACT7 Shared Task: ASR Consistency Filtering and Speaker-Adaptive Post-Processing for Arabic Speech Diacritization](https://aclanthology.org/2026.osact-1.32/) (Abu Horaira & Chowdhury, OSACT 2026)
ACL