ARCADE: A City-Scale Corpus for Fine-Grained Arabic Dialect Tagging

TL;DR

ARCADE tackles the need for city-level Arabic dialect tagging by building a radio-based corpus with fine-grained geographic labels. It employs an automated recording pipeline and a multi-annotator workflow to produce 6,907 annotations over 3,790 clips from 58 cities across 19 countries, labeling emotion, speech type, dialect category, and sample validity. The paper provides detailed data characterization, reliability analyses, and audio-quality metrics, and releases the dataset with documentation and a reproducible collection protocol. This resource enables robust, city-aware dialect identification and multi-task modeling, supporting geolinguistic studies and cross-domain transfer while highlighting practical challenges like data imbalance and broadcast noise.

Abstract

The Arabic language is characterized by a rich tapestry of regional dialects that differ substantially in phonetics and lexicon, reflecting the geographic and cultural diversity of its speakers. Despite the availability of many multi-dialect datasets, mapping speech to fine-grained dialect sources, such as cities, remains underexplored. We present ARCADE (Arabic Radio Corpus for Audio Dialect Evaluation), the first Arabic speech dataset designed explicitly with city-level dialect granularity. The corpus comprises Arabic radio speech collected from streaming services across the Arab world. Our data pipeline captures 30-second segments from verified radio streams, encompassing both Modern Standard Arabic (MSA) and diverse dialectal speech. To ensure reliability, each clip was annotated by one to three native Arabic reviewers who assigned rich metadata, including emotion, speech type, dialect category, and a validity flag for dialect identification tasks. The resulting corpus comprises 6,907 annotations and 3,790 unique audio segments spanning 58 cities across 19 countries. These fine-grained annotations enable robust multi-task learning, serving as a benchmark for city-level dialect tagging. We detail the data collection methodology, assess audio quality, and provide a comprehensive analysis of label distributions. The dataset is available on: https://huggingface.co/datasets/riotu-lab/ARCADE-full

Figures (11)

• Figure 1: The main pipeline for building the ARCADE dataset.
• Figure 2: Screenshot of the custom Gradio-based annotation interface used by native speakers to label the dataset.
• Figure 3: Distribution of audio types across the annotated corpus. The dataset primarily consists of regular single-speaker recordings and multiple-speaker segments, with a smaller subset containing music/no speech or Quranic recitation.
• Figure 4: MSA/dialect breakdown of the collected annotated segments. Dialectal speech forms the largest portion of the data, followed by MSA, then mixed and not applicable categories.
• Figure 5: Emotion distribution (in log scale) in the ARCADE dataset. Emotion annotations were mainly applied to retained segments, as skipped audio clips (typically Quran recitation or music) were not meant to be labeled for emotion. Neutral accounts for the vast majority of annotations, while other emotions are scarce. Note that 42 retained recordings were missing emotion annotations and are excluded from this graph.