NourZohdy/libyan-subregional-did

Sub-Regional Libyan Arabic Speech Corpus and DID Benchmark — release v1.0

Release package for the paper

Nour A. Zohdy and Mansour Essgaer. 2026. Beyond Monolithic Libyan Arabic: A Sub-Regional Speech Corpus and Dialect Identification Benchmark. Proceedings of ArabicNLP 2026.

DOI of this package: https://doi.org/10.5281/zenodo.22802659 Code repository: https://github.com/NourZohdy/libyan-subregional-did

The corpus is Libyan Arabic speech from public YouTube and TikTok media, labeled with one of three regions: Tripolitania (west), Cyrenaica (east), Fezzan (south). It has 133.7 hours of single-speaker segments (3–10 s) from 645 speakers across 161 channels (57 YouTube channels or playlists, 104 TikTok accounts), with speaker-disjoint train/dev/test splits and a second, channel-disjoint split.

We do not redistribute audio. The package gives you the segment manifest, the source identifiers to fetch the audio yourself (a YouTube video ID for 717 of the 739 YouTube recordings; a public account handle plus clip caption for every TikTok clip), and cached embeddings from two pretrained models so the paper's probe benchmarks can be reproduced without any audio.

Files

File	What it is
segments.csv	78,411 rows, one per released segment. Columns below.
data/speaker_disjoint/{train,dev,test}.csv, data/channel_disjoint/{train,dev,test}.csv	The same rows as segments.csv, cut by the split column (paper's main speaker-disjoint benchmark) or the channel_split column, so the dataset viewer shows train/dev/test. Use their segment_id column, not their local row number, to index the .npy files.
channels.csv	161 rows, one per channel: platform, majority region, URL or handle where known, channel split, size.
ecapa_spkrec_192.npy	float32 [78411, 192]. Speaker-recognition ECAPA-TDNN embedding (speechbrain/spkrec-ecapa-voxceleb) of each segment. Row i is segments.csv row i. Model E2 in the paper.
lid_ecapa_256.npy	float32 [78411, 256]. VoxLingua107 language-ID ECAPA embedding (speechbrain/lang-id-voxlingua107-ecapa) of each segment. Row i is segments.csv row i. Model E5 in the paper.
pipeline_v1.0.zip	The libyan_did Python package: discovery, harvesting, diarization and segmentation, quality gates, clustering and linking, validation tooling, split generation, and the benchmark scripts (E0–E6).
SHA256SUMS	Checksums of every file above.
LICENSE-DATA, LICENSE-CODE	CC BY 4.0 for the manifests and embeddings; MIT for the code.

Both embedding caches come from pretrained models only. No embeddings from the fine-tuned model (E6) are released, so no label information enters the released features.

segments.csv columns

Column	Meaning
segment_id	0-based row index; also the row index into both .npy files.
file_id	Recording (one video or clip). 1,281 distinct values.
channel_id	Source channel: a YouTube channel or playlist name, or tiktok_<handle>. 161 distinct values.
platform	youtube or tiktok.
youtube_id	YouTube video ID (YouTube rows). Empty for 22 of the 739 YouTube recordings (1,913 segments) whose ID could not be recovered.
source_url	Channel or playlist URL where we recorded one (14 of the 57 YouTube channels). Locate YouTube audio by youtube_id, not by this column.
tiktok_handle	Public TikTok account handle (TikTok rows).
local_filename	Our local recording filename for TikTok clips; it contains the clip's public caption and lets you match the clip on the account page.
start_time, end_time, duration	Segment boundaries in seconds within the recording, after our diarization and pause-based cutting.
region	Released label: Tripolitania, Cyrenaica, or Fezzan. This is the label after human review.
provenance_region	The channel's region before review. Differs from region for the 31 relabelled speakers (6,285 segments).
speaker_id	Anonymous speaker identifier (spk_00000 …), constant across recordings and channels for the same speaker after cross-episode linking. 645 distinct values.
split	Speaker-disjoint split: train, dev, test (80/10/10 by duration). A speaker appears in exactly one split. This is the paper's main benchmark split.
channel_split	Channel-disjoint split: train, dev, test. A channel appears in exactly one split. Used for the robustness test in Section 5.7 of the paper.
snr_db, speech_ratio, quality	Signal-to-noise estimate, voice-activity speech ratio, and the resulting quality tag (clean = SNR ≥ 15 dB).

Rebuilding the audio

YouTube: download the video by youtube_id (for example with yt-dlp), convert to 16 kHz mono WAV, and cut [start_time, end_time]. TikTok: open the account https://www.tiktok.com/@<tiktok_handle>, find the clip whose caption matches local_filename, and cut the same interval. Segment boundaries are relative to the full recording as published. The pipeline in pipeline_v1.0.zip contains the harvesting and segmentation code we used.

Reproducing the probe benchmarks

import numpy as np, pandas as pd
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import f1_score

seg = pd.read_csv("segments.csv")
X = np.load("lid_ecapa_256.npy")            # or ecapa_spkrec_192.npy
X /= np.linalg.norm(X, axis=1, keepdims=True)
tr, te = seg.split == "train", seg.split == "test"
clf = LogisticRegression(C=1.0, max_iter=5000).fit(X[tr], seg.region[tr])
print(f1_score(seg.region[te], clf.predict(X[te]), average="macro"))   # ≈ 0.66 for LID-ECAPA (E5)

Use seg.channel_split instead of seg.split for the channel-disjoint setting.

Labels and their limits

Labels start from the channel's known region (provenance) and were reviewed by one annotator per speaker cluster; 95.5% of kept speakers kept their provenance region. A second, blind review by a co-author agreed with 81.4% of released labels (κ = 0.72). A segment-level audit found that about 17% of test segments (33% for Fezzan) are Modern Standard Arabic rather than dialect. Treat the labels as reviewed provenance labels, not gold perceptual labels. See the paper's Section 3.3 and Limitations.

Licenses and ethics

Manifests and embeddings: CC BY 4.0 (LICENSE-DATA). Code: MIT (LICENSE-CODE). These licenses do not cover the underlying YouTube or TikTok audio, which stays under its publishers' rights and the platforms' terms.

Speakers are identified only by anonymous IDs. For TikTok, the public account handle is included because it is the only way to locate the clip; for most TikTok accounts, which contain a single speaker, the handle identifies that speaker. If you are a speaker or publisher and want your entries removed, email nour.tammous@idp.sebhau.edu.ly and we will remove them in the next version of this record.

Citation

@inproceedings{zohdy2026libyan,
  title     = {Beyond Monolithic Libyan Arabic: A Sub-Regional Speech Corpus and Dialect Identification Benchmark},
  author    = {Zohdy, Nour A. and Essgaer, Mansour},
  booktitle = {Proceedings of ArabicNLP 2026},
  year      = {2026}
}

Data and code: Zohdy, N. A. and Essgaer, M. (2026). Sub-Regional Libyan Arabic Speech Corpus and Dialect Identification Benchmark: manifests, cached embeddings, and pipeline (v1.0). Zenodo. https://doi.org/10.5281/zenodo.22802659