@inproceedings{zohdy-essgaer-2026-beyond,
title = "Beyond Monolithic {L}ibyan {A}rabic: A Sub-Regional Speech Corpus and Dialect Identification Benchmark",
author = "Zohdy, Nour A. and
Essgaer, Mansour",
editor = "Abdul-Mageed, Muhammad and
Elsayed, Tamer and
Bouamor, Houda and
El-Beltagy, Samhaa and
Al-Khalifa, Hend and
Abdelali, Ahmed and
Elmadany, AbdelRahim and
El Mekki, Abdellah and
Darwish, Kareem and
Alkhereyf, Sakhar and
Luqman, Hamzah and
Khalifa, Salam and
Magdy, Samar and
Alyafeai, Zaid and
Hamed, Injy and
Zaghouani, Wajdi and
Zitouni, Imed",
booktitle = "Proceedings of The Fourth {A}rabic Natural Language Processing Conference",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.arabicnlp-main.33/",
pages = "485--496",
ISBN = "979-8-89176-473-6",
abstract = "Spoken Arabic dialect identification (DID) has advanced quickly, but Libyan Arabic remains under-served and, where it appears, is labeled only at the country level, hiding the variation between the Libyan coast and interior. We present a Libyan Arabic speech corpus annotated at the sub-regional level: Tripolitania (west), Cyrenaica (east), and Fezzan (south). It contains 133.7 hours of single-speaker segments from 645 speakers across 161 channels (YouTube and TikTok sources). An automated pipeline (diarization, speaker embedding, clustering, cross-episode linking) produces silver labels from channel provenance. One annotator then reviewed the resulting speaker clusters and confirmed the provenance region for 95.5{\%} of the kept speakers (a within-annotator figure). We release speaker-disjoint splits. An off-the-shelf Arabic DID model recognizes our verified Libyan speech as Libyan only 74{\%} of the time and cannot resolve the region. On the new three-way task, the encoder matters more than the classifier: a probe over speaker-recognition ECAPA embeddings reaches 0.525 macro-F1, over a language-identification ECAPA reaches 0.660, and fine-tuning that encoder reaches 0.702 (75.5{\%} accuracy), far above the 0.20 floor. A channel-disjoint re-test reaches 0.609, versus 0.702 speaker-disjoint; Fezzan is hardest at 0.313. The task is learnable, but Fezzan performance remains low. We release the corpus, cached embeddings, and code."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zohdy-essgaer-2026-beyond">
<titleInfo>
<title>Beyond Monolithic Libyan Arabic: A Sub-Regional Speech Corpus and Dialect Identification Benchmark</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nour</namePart>
<namePart type="given">A</namePart>
<namePart type="family">Zohdy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mansour</namePart>
<namePart type="family">Essgaer</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>Proceedings of The Fourth Arabic Natural Language Processing Conference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Muhammad</namePart>
<namePart type="family">Abdul-Mageed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tamer</namePart>
<namePart type="family">Elsayed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Houda</namePart>
<namePart type="family">Bouamor</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Samhaa</namePart>
<namePart type="family">El-Beltagy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<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">Ahmed</namePart>
<namePart type="family">Abdelali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">AbdelRahim</namePart>
<namePart type="family">Elmadany</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Abdellah</namePart>
<namePart type="family">El Mekki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kareem</namePart>
<namePart type="family">Darwish</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sakhar</namePart>
<namePart type="family">Alkhereyf</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hamzah</namePart>
<namePart type="family">Luqman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Salam</namePart>
<namePart type="family">Khalifa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Samar</namePart>
<namePart type="family">Magdy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zaid</namePart>
<namePart type="family">Alyafeai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Injy</namePart>
<namePart type="family">Hamed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wajdi</namePart>
<namePart type="family">Zaghouani</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Imed</namePart>
<namePart type="family">Zitouni</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">979-8-89176-473-6</identifier>
</relatedItem>
<abstract>Spoken Arabic dialect identification (DID) has advanced quickly, but Libyan Arabic remains under-served and, where it appears, is labeled only at the country level, hiding the variation between the Libyan coast and interior. We present a Libyan Arabic speech corpus annotated at the sub-regional level: Tripolitania (west), Cyrenaica (east), and Fezzan (south). It contains 133.7 hours of single-speaker segments from 645 speakers across 161 channels (YouTube and TikTok sources). An automated pipeline (diarization, speaker embedding, clustering, cross-episode linking) produces silver labels from channel provenance. One annotator then reviewed the resulting speaker clusters and confirmed the provenance region for 95.5% of the kept speakers (a within-annotator figure). We release speaker-disjoint splits. An off-the-shelf Arabic DID model recognizes our verified Libyan speech as Libyan only 74% of the time and cannot resolve the region. On the new three-way task, the encoder matters more than the classifier: a probe over speaker-recognition ECAPA embeddings reaches 0.525 macro-F1, over a language-identification ECAPA reaches 0.660, and fine-tuning that encoder reaches 0.702 (75.5% accuracy), far above the 0.20 floor. A channel-disjoint re-test reaches 0.609, versus 0.702 speaker-disjoint; Fezzan is hardest at 0.313. The task is learnable, but Fezzan performance remains low. We release the corpus, cached embeddings, and code.</abstract>
<identifier type="citekey">zohdy-essgaer-2026-beyond</identifier>
<location>
<url>https://aclanthology.org/2026.arabicnlp-main.33/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>485</start>
<end>496</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Beyond Monolithic Libyan Arabic: A Sub-Regional Speech Corpus and Dialect Identification Benchmark
%A Zohdy, Nour A.
%A Essgaer, Mansour
%Y Abdul-Mageed, Muhammad
%Y Elsayed, Tamer
%Y Bouamor, Houda
%Y El-Beltagy, Samhaa
%Y Al-Khalifa, Hend
%Y Abdelali, Ahmed
%Y Elmadany, AbdelRahim
%Y El Mekki, Abdellah
%Y Darwish, Kareem
%Y Alkhereyf, Sakhar
%Y Luqman, Hamzah
%Y Khalifa, Salam
%Y Magdy, Samar
%Y Alyafeai, Zaid
%Y Hamed, Injy
%Y Zaghouani, Wajdi
%Y Zitouni, Imed
%S Proceedings of The Fourth Arabic Natural Language Processing Conference
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-473-6
%F zohdy-essgaer-2026-beyond
%X Spoken Arabic dialect identification (DID) has advanced quickly, but Libyan Arabic remains under-served and, where it appears, is labeled only at the country level, hiding the variation between the Libyan coast and interior. We present a Libyan Arabic speech corpus annotated at the sub-regional level: Tripolitania (west), Cyrenaica (east), and Fezzan (south). It contains 133.7 hours of single-speaker segments from 645 speakers across 161 channels (YouTube and TikTok sources). An automated pipeline (diarization, speaker embedding, clustering, cross-episode linking) produces silver labels from channel provenance. One annotator then reviewed the resulting speaker clusters and confirmed the provenance region for 95.5% of the kept speakers (a within-annotator figure). We release speaker-disjoint splits. An off-the-shelf Arabic DID model recognizes our verified Libyan speech as Libyan only 74% of the time and cannot resolve the region. On the new three-way task, the encoder matters more than the classifier: a probe over speaker-recognition ECAPA embeddings reaches 0.525 macro-F1, over a language-identification ECAPA reaches 0.660, and fine-tuning that encoder reaches 0.702 (75.5% accuracy), far above the 0.20 floor. A channel-disjoint re-test reaches 0.609, versus 0.702 speaker-disjoint; Fezzan is hardest at 0.313. The task is learnable, but Fezzan performance remains low. We release the corpus, cached embeddings, and code.
%U https://aclanthology.org/2026.arabicnlp-main.33/
%P 485-496
Markdown (Informal)
[Beyond Monolithic Libyan Arabic: A Sub-Regional Speech Corpus and Dialect Identification Benchmark](https://aclanthology.org/2026.arabicnlp-main.33/) (Zohdy & Essgaer, ArabicNLP 2026)
ACL