Dialectal Arabic Lahgtna v2

Large-scale multi-dialect Arabic speech dataset — 3,000+ hours across 13 Arabic dialects — for training and evaluating dialectal Arabic ASR systems. Part of the Lahgtna (لهجتنا) project for dialect-aware Arabic speech AI.

Dataset Summary

• ~611K utterances / 3,000+ hours of transcribed dialectal Arabic speech
• **13 Arabic dialects **, labeled per utterance
• 16 kHz mono audio
• Transcripts written in authentic dialectal orthography (not normalized to MSA), with partial diacritization in many samples

Split	Rows
train	608,100
test	3,100

Supported Tasks

• Automatic Speech Recognition (ASR) — dialectal Arabic transcription
• Spoken Dialect Identification — via the language column
• TTS / speech data pipelines — transcribed dialectal speech at 16 kHz

Dataset Structure

Each row contains:

Column	Type	Description
audio	Audio (16 kHz)	Speech segment
transcript_text	string	Dialectal Arabic transcript
language	string	Dialect code (see below)

Dialect Codes

Code	Language / Dialect	Region
bh	Bahraini	Bahrain
dz	Algerian	Algeria
eg	Egyptian	Egypt
iq	Iraqi	Iraq
lb	Lebanese	Lebanon
ly	Libyan	Libya
ma	Moroccan (Darija)	Morocco
ps	Palestinian	Palestine
sa	Saudi	Saudi Arabia
sd	Sudanese	Sudan
sy	Syrian	Syria
tn	Tunisian	Tunisia

Usage

from datasets import load_dataset

ds = load_dataset("oddadmix/dialectal-arabic-lahgtna-v2", split="train", streaming=True)

sample = next(iter(ds))
print(sample["language"], "-", sample["transcript_text"])

The train split is ~337 GB in parquet — streaming=True is recommended unless you need the full dataset on disk.

Filter to a single dialect:

eg_only = ds.filter(lambda x: x["language"] == "eg")

Transcription Notes

• Transcripts preserve dialectal spelling and vocabulary rather than converting to MSA
• Many samples include partial diacritics (تشكيل) to capture dialectal pronunciation
• Non-speech events (e.g. [exhale]) are annotated inline
• Code-switching into English/French is transcribed as spoken (in Arabic script where uttered that way)

Intended Use

Built for fine-tuning ASR models on dialectal Arabic — e.g. Whisper, omniASR CTC, and Qwen3-ASR variants — and as training/eval data for spoken dialect identification. The test split is held out for evaluation.

Limitations

• Dialect labels are per-utterance; intra-dialect regional variation is not sub-labeled
• Dialect distribution is not balanced across the 13 labels
• Transcription conventions for dialectal Arabic are not fully standardized; minor orthographic inconsistency is expected

Citation

@misc{lahgtna_dialectal_arabic_v2_2026,
  author = {oddadmix},
  title = {Dialectal Arabic Lahgtna v2: A Multi-Dialect Arabic Speech Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/oddadmix/dialectal-arabic-lahgtna-v2}
}