Paper deep dive
Learning the Arabic Dialect Continuum as a Continuous Space: A Regression Approach to Speaker Origin Prediction
Mohamed Aziz Khadraoui, Adel Ammar, Bilel Benjdira, Zahid Khan, Skander Turki, Wadii Boulila
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/23/2026, 2:35:56 AM
Summary
This paper proposes a regression-based approach for Arabic dialect geolocation, modeling dialectal variation as a continuous geographic space rather than discrete categories. The authors utilize a hierarchical neural architecture fusing XLS-R-300M and Whisper-large-v3 encoder representations with phonotactic descriptors to predict speaker origin as continuous latitude-longitude coordinates. Using a spherical geodesic loss, the model achieves a median localization error of 481.2 km on the ARCADE corpus, with auxiliary country and city classification accuracies of 64.5% and 45.2%, respectively. The study validates the Arabic dialect continuum hypothesis and demonstrates the model's generalization capabilities through zero-shot city-masking protocols.
Entities (8)
Relation Signals (7)
Regression-based approach → evaluatedon → Arcade
confidence 98% · evaluated under a leakage-free 5-fold GroupKFold protocol... on the ARCADE corpus
Regression-based approach → optimizeswith → Spherical geodesic loss
confidence 95% · A spherical geodesic loss directly optimizes great-circle distance on Earth's surface
Regression-based approach → uses → Whisper large-v3
confidence 95% · fuses frame-level XLS-R-300M and Whisper-large-v3 encoder representations
Regression-based approach → uses → XLS-R-300M
confidence 95% · fuses frame-level XLS-R-300M and Whisper-large-v3 encoder representations
Regression-based approach → supports → Arabic Dialect Continuum
confidence 90% · provides quantitative support for the Arabic dialect continuum hypothesis
Regression-based approach → uses → GroupKFold
confidence 85% · Under a leakage-free 5-fold GroupKFold protocol grouped by source recording
Regression-based approach → uses → Mantel test
confidence 85% · A permutation Mantel test on the learned latent space provides quantitative support
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We present a regression-based approach to Arabic dialect geolocation that models dialectal variation as a continuous geographic space rather than discrete categories. Speaker origin is predicted as continuous latitude-longitude coordinates using a hierarchical neural architecture that fuses frame-level XLS-R-300M and Whisper-large-v3 encoder representations with phonotactic descriptors through a Transformer encoder and a learnable attention-pooled query. A spherical geodesic loss directly optimizes great-circle distance on Earth's surface, avoiding distortions inherent to planar coordinate regression. Under a leakage-free 5-fold GroupKFold protocol grouped by source recording, our model attains a pooled median localization error of 481.2 km. Auxiliary country and city heads reach 64.5% and 45.2% accuracy, respectively. A permutation Mantel test on the learned latent space provides quantitative support for the Arabic dialect continuum hypothesis. To probe true generalization, we further introduce a city-masking protocol in which two cities per fold are removed from training but retained in validation. Under this zero-shot regime, the mean error rises to 1173.3 km, a 1.32x degradation relative to seen cities. Our findings establish continuous geographic modeling as a principled framework for Arabic dialect geolocation and quantify both its strengths and the substantial headroom that remains.
Tags
Links
- Source: https://arxiv.org/abs/2607.19751v1
- Canonical: https://arxiv.org/abs/2607.19751v1
Trouble viewing inline? Open PDF directly →
Full Text
77,207 characters extracted from source content.
Expand or collapse full text
v 2026 Paper Learning the Arabic Dialect Continuum as a Continuous Space: A Regression Approach to Speaker Origin Prediction Mohamed Aziz Khadraoui1 Adel Ammar2 Corresponding author: Bilel Benjdira2 Zahid Khan2 Skander Turki2 Wadii Boulila2 Higher School of Communication of Tunis (SUP’COM), Tunisia Robotics and Internet of Things Laboratory, Prince Sultan University, Riyadh, Saudi Arabia Abstract We present a regression-based approach to Arabic dialect geolocation that models dialectal variation as a continuous geographic space rather than discrete categories. Speaker origin is predicted as continuous latitude–longitude coordinates using a hierarchical neural architecture that fuses frame-level XLS-R-300M and Whisper-large-v3 encoder representations with phonotactic descriptors through a Transformer encoder and a learnable attention-pooled query. A spherical geodesic loss directly optimizes great-circle distance on Earth’s surface, avoiding distortions inherent to planar coordinate regression. Under a leakage-free 5-fold GroupKFold protocol grouped by source recording, our model attains a pooled median localization error of 481.2 km. Auxiliary country and city heads reach 64.5% and 45.2% accuracy, respectively. A permutation Mantel test on the learned latent space provides quantitative support for the Arabic dialect continuum hypothesis (r=0.131r=0.131, p≈10−4p≈ 10^-4 over 10,000 permutations). To probe true generalization, we further introduce a city-masking protocol in which two cities per fold are removed from training but retained in validation; under this zero-shot regime mean error rises to 1173.3 km, a 1.32×1.32× degradation relative to seen cities. Our findings establish continuous geographic modeling as a principled framework for Arabic dialect geolocation and quantify, both its strengths and the substantial headroom that remains. 1 Introduction Arabic dialectal variation spans a vast geographic region from Morocco to Oman, encompassing hundreds of millions of speakers across 22 countries. While Modern Standard Arabic (MSA) serves as a formal written standard, spoken communication predominantly occurs in regional dialects that differ substantially in phonology, morphology, lexicon, and syntax. These differences are not merely surface-level: mutual intelligibility between geographically distant varieties (for example, Moroccan Darija and Gulf Arabic) can be severely limited, while neighbouring varieties shade into one another almost imperceptibly. Understanding and automatically identifying these dialectal differences has profound implications for speech recognition, machine translation, information retrieval, voice assistants, forensic speaker profiling, and sociolinguistic research. The spoken modality is central to this problem. Dialectal distinctions are realized most richly in speech, through phonetic, prosodic, and rhythmic cues that are largely lost when utterances are transcribed into a non-standardized orthography. Consequently, while a large body of work has addressed dialect identification from text, identifying dialect (and, more ambitiously, speaker origin) directly from the audio signal is both more faithful to how dialects are used and considerably more challenging, owing to channel variability, background noise, code-switching with MSA or other languages, and the scarcity of geographically labelled speech corpora. Early speech-based systems framed the problem as a small, coarse classification task (typically distinguishing four or five regional groups) and only recently has city-level granularity become feasible with the release of large, metadata-rich speech corpora. Traditional approaches to Arabic dialect identification, whether text- or speech-based, treat dialects as discrete, mutually exclusive categories corresponding to countries, regions, or cities. However, this discretization is in tension with established sociolinguistic theory, which characterizes dialectal variation as existing along continuous gradients (dialect continua) rather than exhibiting sharp boundaries Chambers and Trudgill (1998); Heeringa and Nerbonne (2001); Nerbonne (2009). Speakers near geographic borders often exhibit mixed dialectal features, and migration patterns create complex, overlapping isoglosses that defy simple categorization. Imposing hard categorical boundaries on this gradient structure discards precisely the information that distinguishes a principled geographic model from an arbitrary partition of space. We propose a paradigm shift from discrete classification to continuous regression, formulating dialect geolocation as the prediction of precise geographic coordinates on Earth’s surface. Conceptually, this reframes Arabic dialect identification as a speech-based geolocation problem, analogous to image geolocation in computer vision, where the goal is to place an input on a map rather than into a bin. This regression-based approach offers several advantages: (1) it naturally captures gradient variation without imposing artificial decision boundaries, (2) it enables fine-grained predictions at unseen locations, (3) it reduces sensitivity to class imbalance arising from uneven data collection, and (4) it provides a principled framework for modeling the spatial structure of linguistic variation. Crucially, it also yields an interpretable, distance-based error metric (kilometers on the Earth’s surface) that is directly comparable across studies, unlike accuracy figures that depend on an arbitrary choice of dialect classes. The main contributions of this work are as follows: • We formulate Arabic dialect geolocation as a continuous geographic regression task, replacing conventional discrete dialect classification with the prediction of latitude-longitude coordinates on the Earth’s surface. • We propose a hierarchical multi-task architecture that combines spherical geodesic regression with auxiliary country and city classification objectives, enabling the model to jointly capture continuous and discrete aspects of dialectal variation. • We introduce a rigorous evaluation protocol based on the ARCADE corpus Nacar et al. (2026), including geographically informed cross-validation with masked-city splits and geodesic distance metrics, providing a reproducible benchmark for continuous dialect geolocation. • We provide an extensive empirical and linguistic analysis demonstrating that Arabic dialect variation is more accurately modeled as a continuous geographic continuum than as a collection of isolated dialect classes, yielding new insights into the relationship between geographic proximity and speech variation. Experimental results on 2,329 Arabic speech samples spanning 19 countries and 46 cities, evaluated under a leakage-free 5-fold GroupKFold protocol with city-masking for zero-shot probing, demonstrate a pooled median localization error of 481.2 km and country accuracy of 64.5%. Beyond raw accuracy, we analyze model calibration, selective prediction, border-versus-interior behavior, and unseen-city generalization, establishing both practical pathways for deployment and the gap that remains between closed-set evaluation and true geographic generalization. The remainder of this paper is organized as follows. Section 2 reviews the relevant literature on Arabic dialect identification, continuous dialect modeling, and speech-based geolocation. Section 3 formulates the task as a continuous geographic regression problem. Section 4 describes the dataset, data-quality procedures, and preprocessing pipeline. Section 5 presents the proposed methodology, including the model architecture, training objectives, and optimization strategy. Section 6 details the experimental setup and evaluation protocol, while Section 7 presents and discusses the experimental results. Section 8 discusses the limitations of the proposed approach. Finally, Section 9 concludes the paper and outlines directions for future research. 2 Background and Related Works 2.1 Early Dialect Identification Systems Early Arabic dialect identification systems treated dialects as discrete categories. Zaidan et al. Zaidan and Callison-Burch (2014) established the foundation with the Arabic Online Commentary dataset containing 52 million words, while Bouamor et al. Bouamor et al. (2018) created the comprehensive MADAR corpus with parallel translations across 25 city dialects. Traditional approaches using Support Vector Machines and n-gram features Elfardy and Diab (2013) achieved reasonable performance but failed to capture the continuous nature of dialectal variation. We note that all of these foundational systems operate on text; the spoken modality, which is the focus of our work, has historically received far less attention. 2.2 Speech-Based Dialect Identification In contrast to the large text-based literature, spoken Arabic dialect identification (ADI) has matured more slowly, gated by the availability of labelled speech. Ali et al. Ali et al. (2016) introduced automatic dialect detection from Arabic broadcast speech, combining acoustic and lexical (phone- and word-level) features over five regional varieties, and Khurana et al. Khurana et al. (2017) explored multi-view dimensionality reduction for the same task. The ADI17 benchmark of Shon et al. Shon et al. (2020) scaled the problem to 17 country-level dialects using ∼ 3,000 hours of YouTube speech, and end-to-end neural systems built on i-vectors, x-vectors, and convolutional encoders became the dominant paradigm. More recently, self-supervised speech encoders such as wav2vec 2.0 and its multilingual variant XLS-R Babu et al. (2022), together with weakly supervised models such as Whisper Radford et al. (2023), have provided strong transferable representations that we exploit in this work. Parallel efforts have expanded the available labelled speech: the Casablanca corpus Talafha and others (2024), for instance, provides multi-dialectal Arabic speech with transcriptions and dialect labels across several countries. Critically, however, all of these speech systems remain classification systems over a small, fixed set of regions or countries; none predicts continuous speaker origin, and city-level granularity has only become tractable with the recent release of the metadata-rich ARCADE radio corpus Nacar et al. (2026) on which we build. 2.3 Continuous Dialect Modeling The revolutionary shift came with Keleg et al. Keleg et al. (2023), who introduced regression-based modeling that quantifies dialectness continuously. Their AOC-ALDi dataset contains 127,835 sentences with continuous annotations, achieving RMSE values of 0.1403. Building on this, Shaban et al. Sha’ban and Habash (2025) added the Arabic Generality Score dimension, employing etymology-aware edit distance and CAMeL-BERT regression. This two-dimensional continuous space (ALDi × AGS) provides unprecedented granularity in modeling dialectal relationships. 2.4 Transformer-Based Architectures The transformer revolution has profoundly impacted Arabic dialect processing. AraBERT Antoun et al. (2020) pioneered transformer-based Arabic understanding, while MARBERTv2 Abdul-Mageed et al. (2021), trained on 6 billion tweets, achieves 65% accuracy on 18-dialect classification. Recent advances in the NADI shared tasks Abdul-Mageed et al. (2024) demonstrate the effectiveness of ensemble regression methods combining BERT-based models with contrastive learning. Elaraby et al. Elaraby and Abdul-Mageed (2018) showed that BiLSTM architectures achieve 92.65% accuracy, while parameter-efficient approaches demonstrate comparable performance using only 2.5% of model parameters. 2.5 Key Datasets and Resources The MADAR corpus Bouamor et al. (2018) remains the most comprehensive resource with 47,466 dialectal words mapped to 1,045 concepts across 25 city dialects. QADI Abdelali et al. (2021) contributes 540,590 tweets from 18 Arab countries with 91.5% intrinsic accuracy. The Arabic Online Commentary dataset Zaidan and Callison-Burch (2011) provides foundational coverage of Egyptian, Gulf, and Levantine dialects. For continuous modeling, the AOC-ALDi corpus with continuous dialectness annotations and the NADI 2024 dataset Abdul-Mageed et al. (2024) featuring multi-label geographic annotations are essential. These resources enable fine-grained analysis at unprecedented granularity, supporting regression-based approaches to speaker origin prediction. 2.6 Challenges and Open Problems Code-switching between Modern Standard Arabic and dialects presents fundamental challenges, with word error rates of 28–54% for code-switched speech Hamed and others (2025). The lack of standardized orthography, addressed partially by CODA Habash et al. (2012), continues to impact model generalization. Data scarcity affects numerous dialects disproportionately. In fact, while Egyptian Arabic has extensive resources, rural North African variants remain severely underrepresented. Evaluation metrics for continuous predictions remain inadequate. Traditional metrics like accuracy fail to capture the nuanced nature of continuous dialectness, while regression metrics like RMSE don’t fully represent linguistic validity. The absence of standardized evaluation frameworks hinders reproducible research. 2.7 Tools and Frameworks The tools most widely used for Arabic dialect work operate on text rather than speech. CAMeL Tools Obeid and others (2020) provides comprehensive Arabic text NLP capabilities including text-based dialect identification for 25 dialects. MADAMIRA Pasha and others (2014) offers sophisticated textual morphological analysis and disambiguation. AraVec Soliman and others (2017) provides pre-trained word embeddings learned from written corpora. These text-oriented resources underpin most modern Arabic dialect systems but are not directly applicable to the raw audio signal; speech-based systems instead rely on acoustic toolkits (e.g. Kaldi-style i-vector/x-vector pipelines) and, increasingly, on self-supervised speech encoders such as XLS-R Babu et al. (2022) and Whisper Radford et al. (2023), as adopted in this work. 2.8 Comparison with Prior Work Table 1 situates our approach against representative prior work along the dimensions most relevant to this paper: the input modality (text vs. speech), the task formulation, the geographic granularity of the labels, and the form of the output. The table makes explicit that prior speech-based systems are classifiers over a handful of regions or countries, prior continuous-modeling work (ALDi/AGS) operates on text and predicts a dialectness score rather than a location, and our work is the only entry that predicts continuous geographic coordinates directly from speech. Table 1: Comparison of representative Arabic dialect identification work. “Granularity” refers to the geographic resolution of the labels; “Output” to what the system predicts. Our work is the only one combining the speech modality with continuous geographic regression. Work Modality Task Granularity Output Zaidan and Callison-Burch (2014) Text Classification 4 regions Dialect label Bouamor et al. (2018) Text Classification 25 cities City label Abdul-Mageed et al. (2024) Text Classification 18–21 countries Country/province label ALDi Keleg et al. (2023) Text Regression — Dialectness score ∈[0,1]∈[0,1] Ali et al. (2016) Speech Classification 5 regions Dialect label ADI17 Shon et al. (2020) Speech Classification 17 countries Country label ARCADE Nacar et al. (2026) Speech Classification 58 cities City/dialect tag This work Speech Regression Coordinates (lat, lon) on the sphere 2.9 Motivation for This Work Despite significant progress, existing work has two critical limitations. First, most systems continue to frame dialect identification as discrete classification, imposing categorical boundaries that contradict sociolinguistic evidence of continuous variation. Second, geographic-linguistic modeling incorporating spatial relationships between dialect communities remains underexplored for speaker origin prediction. Our work addresses these gaps by developing a continuous regression framework grounded in spherical geometry, providing both improved localization accuracy and theoretical insights into the structure of Arabic dialectal variation. 3 Problem Formulation In this section, we formally define the task of fine-grained Arabic dialect geolocation. Unlike traditional approaches that treat dialect identification as a discrete classification problem over broad and often arbitrary regions, we frame the objective as a continuous geographic regression task aimed at directly predicting a speaker’s spatial coordinates. To robustly model this continuous space, we map standard latitude and longitude representations into a 3D spherical coordinate system, mitigating inherent planar distortions and boundary discontinuities. Finally, recognizing that human geography is intrinsically structured, we establish a hierarchical multi-task learning framework. This framework jointly optimizes continuous coordinate regression alongside categorical country and city classification, thereby enriching the supervisory signal and ensuring geographically consistent predictions. 3.1 Dialect Recognition as a Continuous Geographic Regression Task We formulate Arabic dialect geolocation as predicting continuous geographic coordinates from speech input. Given an audio signal x, the task is to estimate the speaker’s origin as latitude-longitude coordinates (ϕ,λ)(φ,λ) where ϕ∈[−90∘,90∘]φ∈[-90 ,90 ] and λ∈[−180∘,180∘]λ∈[-180 ,180 ]. Traditional dialect classification approaches discretize this space into K predefined regions R1,…,RK\R_1,…,R_K\ and optimize: k^=argmaxk∈1,…,KP(Rk∣). k= _k∈\1,…,K\P(R_k ). (1) The predicted location is then the centroid of region Rk^R_ k. This discretization introduces several problems: (1) artificial boundaries create discontinuities in prediction, (2) within-region variation is ignored, (3) predictions at unseen locations require defining new categories, and (4) class imbalance from uneven geographic sampling creates optimization difficulties. In contrast, our regression approach directly predicts geographic coordinates: (ϕ^,λ^)=fθ(),( φ, λ)=f_θ(x), (2) where fθf_θ is a neural network with parameters θ. This formulation naturally handles gradient variation, enables interpolation to unseen locations, and provides fine-grained localization without categorical constraints. 3.2 Spherical Geometry and Coordinate Representation Operating directly on latitude-longitude pairs introduces several challenges. First, longitude is periodic with a discontinuity at ±180∘± 180 . Second, distance metrics in (ϕ,λ)(φ,λ) space do not correspond to actual geographic distance: one degree of longitude at the equator spans ∼ 111 km, while at 60∘60 latitude it spans only ∼ 56 km. Third, mean squared error in coordinate space distorts spatial relationships near the poles. While these issues are less critical when the scope is restricted solely to the Arab world region, given its limited latitude and longitude ranges, we aim for a generalized approach that remains robust across broader geographic contexts. We address these issues by representing locations as unit vectors on the sphere. Each coordinate pair (ϕ,λ)(φ,λ) is mapped to 3D Cartesian coordinates: x=cos(ϕ)cos(λ)y=cos(ϕ)sin(λ)z=sin(ϕ) \ aligned x&= (φ) (λ)\\ y&= (φ) (λ)\\ z&= (φ) aligned . (3) yielding a unit vector =(x,y,z)y=(x,y,z) with ‖=1\|y\|=1. In this representation, the great-circle distance between two points is simply: d(1,2)=R⋅arccos(1⋅2),d(y_1,y_2)=R· (y_1·y_2), (4) where R=6371R=6371 km is Earth’s mean radius. This formulation eliminates discontinuities, ensures distance computation along Earth’s surface, and simplifies loss function design. 3.3 Hierarchical Geographic Structure Geographic variation exhibits natural hierarchy: countries contain regions, regions contain cities, and cities contain neighborhoods. We exploit this structure through multi-task learning, jointly predicting: • Fine-grained coordinates (ϕ,λ)(φ,λ) via regression. • Country label c∈1,…,Cc∈\1,…,C\ via classification. • City label s∈1,…,Ss∈\1,…,S\ via classification. The hierarchical tasks serve multiple purposes: (1) auxiliary classification objectives provide additional supervision signal, (2) predicted country/city labels can condition coordinate prediction, improving accuracy, (3) classification confidence scores enable uncertainty quantification, and (4) hierarchical predictions offer graceful degradation when fine-grained localization fails. 4 Dataset 4.1 Data Sources Our work builds on the ARCADE corpus Nacar et al. (2026), a city-scale collection of Arabic radio speech segmented into short clips, each annotated with metadata including the broadcasting station, its city, and its country; we refer the reader to the ARCADE paper for the full description of the streaming-based collection, segmentation, and annotation pipeline. From this collection we derive our working set through the data-quality filtering described below. After filtering, our dataset comprises 2,329 radio recordings across 19 countries and 46 cities. The retained recordings span dialectal regions including Maghrebi, Egyptian, Levantine, Iraqi, Gulf, Sudanese, and Yemeni varieties. For each recording we attach a precise latitude–longitude coordinate obtained by geocoding the (city, country) pair against the Nominatim geocoding service OpenStreetMap Foundation (2024). 4.2 Data Collection and Quality Control The audio was not collected by us through elicitation; it consists of naturalistic radio-broadcast speech harvested from streaming services across the Arab world, as detailed in the ARCADE paper Nacar et al. (2026). Our contribution at the data level is the quality filtering applied on top of that collection to produce a clean, geographically consistent subset suitable for coordinate regression. Concretely, starting from the ARCADE clips we (1) discard cities with fewer than five samples (folded into an __OTHER__ bucket during training, see Section 5.3.3), (2) remove geographic outliers whose geocoded coordinates fall more than 20002000 km from their country centroid (typically the result of ambiguous or mis-resolved city names), and (3) drop clips with missing or unresolvable city/country metadata. This filtering reduces the raw collection to the 2,329 recordings used throughout this paper. We deliberately do not re-describe the upstream recording-quality and dialect-validity checks, which are the responsibility of the ARCADE pipeline and are documented in Nacar et al. (2026). 4.3 Geographic Distribution The dataset exhibits substantial geographic diversity with significant sampling-density variations. Algeria contributes the largest share of the validation pool (507 samples), followed by Sudan (290), Egypt (250), Morocco (229), and the United Arab Emirates (212). At the city level, El Obeid (171), Cairo (152), Algiers (121), Amman (103), and Fes (101) are the densest. Several countries (Qatar, Bahrain, Oman) are represented by fewer than 50 samples, and several cities by only a handful, reflecting both population demographics and practical data-collection constraints. To avoid overfitting to over-represented regions, this imbalance necessitates a careful evaluation methodology, such as our leakage-free GroupKFold split by source recording (Section 5.3.3). Figure 1 maps this distribution, with each city sized by sample count and coloured by mean recording SNR, making both the geographic spread and the per-location quality variation visible at a glance. Figure 1: Geographic distribution of the 2,329 audio samples across 19 countries and 46 cities. Each circle marks a city: its size is proportional to the number of samples collected from that location, while its color encodes the mean signal-to-noise ratio (SNR, in dB) of the recordings at that city, as shown by the color bar (red ≈0≈ 0 dB indicates noisier audio, green ≈30≈ 30 dB cleaner audio). The colour thus summarizes recording quality per location and is independent of the sample-count encoding given by circle size. 4.4 Annotation and Preprocessing Because the source material is radio broadcasts, we do not have access to per-speaker self-reported origins. Instead, the ground-truth location of each clip is taken to be the city of its broadcasting station, as recorded in the ARCADE metadata Nacar et al. (2026); we performed a manual review of the station-to-city assignments and discarded clips with inconsistent or ambiguous geographic metadata. Each (city, country) pair is then mapped to a coordinate by geocoding against the Nominatim service OpenStreetMap Foundation (2024), yielding the city-centroid latitude–longitude used as the regression target. This station-city assumption is a deliberate simplification: it treats the dialect of a station’s broadcast as representative of its city, which is reasonable for regional stations but introduces label noise for national broadcasters whose presenters may originate elsewhere. All audio is resampled to 16 kHz mono. We further enforce that no source recording appears in both the training and validation partitions of any fold (Section 5.3.3), so that evaluation reflects geographic generalization rather than recording-level memorization. 4.5 Data Statistics Table 2 summarizes the key properties of the filtered dataset and the parameters of our data-quality pipeline. It records the post-filtering corpus size (2,329 samples spanning 19 countries and 46 cities), the most heavily sampled country (Algeria, 507 samples) and city (El Obeid, 171 samples), which together quantify the geographic imbalance discussed above. The remaining rows document the two thresholds that define our filtering and per-fold label handling — the rare-city threshold (<5<5 training samples in a fold are merged into the __OTHER__ bucket) and the outlier threshold (>2000>2000 km from the country centroid) — as well as the geocoding source (Nominatim) used to obtain coordinates. These last entries make the dataset reproducible from the raw ARCADE collection. Table 2: Key dataset statistics after data-quality filtering. Statistic Value Total samples (post-DQ) 2,329 Countries 19 Cities 46 Largest country (Algeria) 507 samples Largest city (El Obeid) 171 samples Rare-city threshold (per fold) <5<5 training samples → __OTHER__ Geocoding source Nominatim (city, country) lookup DQ outlier threshold >2000>2000 km from country centroid The dataset represents one of the largest collections of geographically annotated Arabic speech samples, enabling robust evaluation of continuous geolocation approaches. 5 Proposed Methodology We address Arabic dialect geolocation as continuous localization over the Earth’s surface rather than classification into predefined geographic cells. Discretization imposes artificial decision boundaries, is sensitive to class imbalance, and cannot predict at locations unseen during training. These limitations are especially acute for Arabic, whose dialect boundaries are fluid, overlapping, and shaped by migration, trade routes, and sociolinguistic contact rather than by administrative borders. The extent to which such variation is in fact spatially continuous is treated here as a hypothesis to be tested rather than assumed, and is examined directly in Section 7.3.1. A continuous formulation alone, however, is not sufficient. Regressing a single coordinate implicitly assumes that each utterance corresponds to one point, whereas a dialectal cue is typically compatible with an extended region and, for shared varieties such as Gulf Arabic, with several mutually distant candidates. Under such ambiguity a point estimator minimizes its expected error by predicting the centroid of the plausible set—a location that may itself be implausible. We therefore model the target as a probability density over the unit sphere and decode its dominant mode, retaining a continuous output space while allowing the model to express multimodal geographic uncertainty instead of averaging it away. Our approach integrates three components: (1) a multi-level acoustic representation that captures complementary dialectal cues; (2) a two-stage architecture that decouples frozen pretrained encoders from a trained fusion backbone; and (3) a multi-task objective coupling a von Mises–Fisher mixture density with a contrastive location-retrieval head and auxiliary coarse-to-fine classification, so that the shared representation is organized geographically at several scales. 5.1 Model Architecture Figure 2: Overview of the proposed Arabic dialect geolocation architecture (GeoArc-F). Complementary frame-level features are extracted once from raw speech by three frozen encoders—XLS-R Babu et al. (2022), the Whisper encoder Radford et al. (2023), and handcrafted phonological descriptors—and cached, decoupling the pretrained front-end from optimisation. A trained backbone projects the three streams into a unified latent space, augments them with positional and modality embeddings, and fuses them through a Transformer encoder followed by query-token attention pooling. The resulting shared 512-d representation is jointly optimised by two geodesic heads—a von Mises–Fisher mixture density over the unit sphere, decoded by its dominant mode, and a contrastive projection aligned to a random-Fourier location encoder Vivanco Cepeda et al. (2023) and decoded by retrieval—together with auxiliary country, region, and city classification heads and an MSA gate that down-weights the geographic loss on dialect-neutral speech. Figure 2 presents a high-level overview of the proposed architecture and its main computational blocks. The design follows a two-stage pipeline. In the first stage, three complementary frozen encoders are evaluated once over the corpus and their frame-level outputs cached, so that the pretrained front-end supplies fixed representations rather than being optimised jointly. This decoupling is deliberate: with only 2,329 dialectal clips, fine-tuning a 300M-parameter encoder is prone to overfitting, and caching renders the five-fold masked-city cross-validation protocol computationally tractable while isolating the contribution of the head design. In the second stage, a trained backbone projects the three streams into a unified latent space, fuses them with a Transformer encoder over the concatenated token sequence, and pools them into a single utterance-level representation. This shared representation is read by parallel prediction heads: two geodesic heads that jointly produce the coordinate, and auxiliary classification heads that shape the latent space during training without contributing to inference. The remainder of this section details each component in isolation, emphasising its role in capturing complementary dialectal cues and in modelling the regional—rather than point-like—nature of dialect geography. 5.1.1 Audio Preprocessing and Feature Extraction All audio signals are resampled to 16 kHz mono and centred to a fixed 10-second window (zero-padded if shorter, centre-cropped if longer) before peak normalization. Dialectal variation manifests across multiple linguistic levels, ranging from fine-grained phonetic realizations to broader prosodic and rhythmic patterns. No single representation captures all these phenomena effectively. We therefore adopt a multi-representation strategy that explicitly targets different aspects of dialectal variation, retaining frame-level (rather than utterance-pooled) features for two of the three modalities so that downstream attention can attend to local cues. Self-Supervised Speech Representations. We extract contextual frame embeddings using XLS-R-300M (facebook/wav2vec2-xls-r-300m). Self-supervised models learn rich phonetic and lexical representations from large volumes of unlabeled multilingual speech, making them particularly suitable for dialectal settings. The native temporal sequence is downsampled by adaptive average pooling to a fixed length of T=32T=32 frames, yielding a tensor in ℝ32×1024R^32× 1024 per utterance. Whisper Encoder Features. To complement self-supervised features we incorporate the encoder of Whisper-large-v3 (openai/whisper-large-v3). Whisper is trained on large-scale multilingual and multi-domain speech data, enabling it to capture language-agnostic acoustic patterns and robust temporal structures. The decoder is discarded to save memory; encoder hidden states are again downsampled to T=32T=32 frames, yielding ℝ32×1280R^32× 1280. Phonotactic and Prosodic Features. While neural embeddings capture high-level abstractions, hand-crafted acoustic features provide complementary cues related to speech rhythm, timbre, and articulation. We compute MFCCs (20 coefficients with first and second derivatives, summarized by per-coefficient mean/std/max/min), spectral descriptors (centroid, bandwidth), zero-crossing rate, and a 64-bin mel spectrogram time-average. The concatenation, padded/truncated to ℝ1000R^1000, forms a single utterance-level vector p. 5.1.2 Feature Projection and Fold-Local Normalization The phonotactic vector p is stored raw in our feature cache. At training time, fold-local normalization statistics (μ,σμ,σ) are computed on the training indices of the current cross-validation fold only, applied at the data-loader level, and finally rescaled by a learnable LayerNorm inside the backbone. This explicit fold-local protocol replaces the global z-score that we identified as a source of evaluation leakage in earlier iterations of this work. Each modality is then projected to a shared d=768d=768 hidden dimension via a per-modality LayerNorm followed by a linear layer. 5.1.3 Frame-Level Backbone with Attention Pool The 32 XLS-R frame tokens, 32 Whisper-encoder frame tokens, and the single phonotactic token are concatenated along the time axis into a sequence of length 2T+1=652T+1=65 tokens. Learned positional embeddings and a learned three-way modality embedding are added to disambiguate the source of each token. The sequence is processed by a 4-layer Transformer encoder (8 heads, dff=4d_f=4d, GELU, pre-norm) which models long-range dependencies across modalities and time. Learnable-Query Attention Pool. A single learnable query token ∈ℝ1×dq ^1× d attends over the entire encoded sequence through a multi-head attention block, yielding the utterance representation pool∈ℝdh_pool ^d. This attention-pool mechanism plays the role of a soft [CLS] token: it lets the network adaptively weight phonetic, prosodic, and self-supervised cues per utterance rather than relying on a fixed mean-pool. 5.1.4 Multi-Task Prediction Heads Geographic variation exhibits a natural hierarchy, progressing from broad regions to fine-grained local variation. To exploit this structure, the model jointly predicts continuous geographic coordinates alongside auxiliary country and city classifications. Geographic Regression Head. Geographic location is predicted as a unit-norm 3D Cartesian vector: geo=regpool‖regpool‖.p_geo= W_reg\,h_pool \|W_reg\,h_pool \|. (5) Unit normalization constrains predictions to Earth’s surface and simplifies distance computation. Country and city heads are simple two-layer MLPs over poolh_pool. Country and City Classification Heads. Auxiliary country and city classifiers provide hierarchical supervision: country=softmax(countrypool),p_country=softmax(W_countryh_pool), (6) city=softmax(citypool).p_city=softmax(W_cityh_pool). (7) These auxiliary tasks regularize the shared representation and encourage geographic consistency. 5.2 Loss Functions and Training Objectives 5.2.1 Spherical Geodesic Loss We optimize angular distance on the unit sphere: ℒangular=1N∑i=1Narccos(clamp(i⋅i,−0.99999,0.99999)).L_angular= 1N _i=1^N (clamp(p_i·y_i,-0.99999,0.99999) ). (8) Angular distance corresponds directly to great-circle distance, aligning the training objective with real-world geographic error. The clamping operation ensures numerical stability near ±1± 1. 5.2.2 Multi-Objective Loss Function The total loss combines regression and classification objectives: ℒtotal=αℒangular+βℒcountry+γℒcity,L_total= _angular+ _country+ _city, (9) with α=0.5α=0.5, β=0.5β=0.5, and γ=0.4γ=0.4. These weights balance fine-grained localization against the two hierarchical classification objectives. Cross-entropy with label smoothing of 0.050.05 is used for both classifiers, and MixUp / SLERP-augmented batches use the corresponding soft targets. 5.3 Training Strategy 5.3.1 Augmentation, EMA and Snapshot Ensembling During training we apply (i) SpecAugment-style time and channel masking on the XLS-R and Whisper frame sequences, (i) MixUp (α=0.2α=0.2) on inputs and on classification targets, with SLERP interpolation on the unit-sphere coordinates so that mixed targets remain valid geographic points, and (i) Gaussian latitude/longitude jitter (σ=0.05∘σ=0.05 ) on the regression targets. Model weights are tracked by an exponential moving average (decay 0.9990.999); each evaluation uses the EMA weights. We further apply test-time augmentation by averaging predictions across five passes with input Gaussian noise (σ=0.02σ=0.02). A snapshot ensemble retains the top-3 EMA snapshots over training, subject to a minimum gap of 10 epochs between retained snapshots; final per-fold predictions average the unit-norm coordinates and the country/city softmaxes from these snapshots. 5.3.2 Optimization and Implementation Details Training employs the configuration shown in Table 3. Models are trained on a single NVIDIA P100 / T4 / A100 GPU (Kaggle pool); a fold typically converges within 40–80 epochs. Table 3: Training hyperparameters and optimization settings. Hyperparameter Setting Optimizer AdamW (β1=0.9 _1=0.9, β2=0.999 _2=0.999, weight decay 5×10−55×10^-5) Learning rate schedule Initial LR 10−410^-4 with cosine annealing to 10−610^-6 Batch size 24 Gradient clipping Maximum norm of 1.0 Max epochs / patience 80 / 18 EMA decay 0.999 MixUp α 0.2 (SLERP on xyzxyz) TTA passes / noise σ 5 / 0.02 Snapshot ensemble top-3, min gap 10 epochs 5.3.3 Leakage-Free Cross-Validation Protocol All evaluation is performed under a 5-fold GroupKFold split where the grouping key is the source recording identifier (extracted from the audio filename). Because the ARCADE corpus is built from segmented radio recordings, multiple short clips can share a parent recording; grouping by source identifier prevents the same recording from contributing to both training and validation. We additionally enforce that all preprocessing statistics (the phonotactic mean/variance and both LabelEncoder vocabularies for country and city) are fit on the training indices of the current fold only. An explicit __OTHER__ bucket at index 0 catches validation labels that fall outside the per-fold training vocabulary, including cities pruned by the per-fold rare-city threshold (<5<5 training samples). City-masking for unseen-city evaluation. Fold 0 serves as a baseline reference (all cities visible during training, modulo the rare-city threshold). For folds 11 through 44, we additionally remove all training samples of two randomly chosen cities; those cities still appear in that fold’s validation set, providing a true zero-shot, unseen-city evaluation. The choice of masked cities is deterministic (seed 20260427+fold_id20260427+fold\_id) and is constrained so that (i) the masked cities appear in both the train and val partitions of the original split, and (i) the post-mask training set retains at least eight unique cities. The resulting per-fold masking is recorded in cv_plan.json and reproduced in Table 5 (column masked). 6 Experimental Evaluation 6.1 Evaluation Metrics 6.1.1 Geodesic Distance Error The primary evaluation metric is the geodesic distance between predicted and ground-truth locations measured along the Earth’s surface. Predictions are represented as unit vectors on the sphere to avoid distortions inherent to planar latitude-longitude regression. Given predicted and ground-truth vectors p and y, the angular distance is computed as: dang=arccos(⋅),d_ang= (p·y), (10) and converted to kilometers by: dgeo=R⋅dang,d_geo=R· d_ang, (11) where R=6371R=6371 km is the mean Earth radius. We report the mean, median, and standard deviation of geodesic error. Median error is emphasized due to the heavy-tailed nature of localization errors, where a small number of extreme failures can disproportionately affect the mean. 6.1.2 Accuracy within Distance Thresholds To facilitate interpretability and comparison with prior work, we report accuracy within predefined distance thresholds: Acc@r=ℙ(dgeo≤r),Acc@r=P(d_geo≤ r), (12) with r∈50,100,250,500,1000,2500r∈\50,100,250,500,1000,2500\ km. These thresholds correspond to intra-city, inter-city, regional, national, multi-country, and continental localization scales, respectively. 6.1.3 Country and City Classification Accuracy For the auxiliary hierarchical heads, we report top-1 accuracy for both country and city prediction. We additionally summarize per-fold accuracy as mean± deviation and the corresponding 95% bootstrap confidence interval over the held-out folds. 6.1.4 Calibration and Confidence Quality We assess probabilistic calibration of the auxiliary classifiers using Expected Calibration Error (ECE) and the Brier score. Let n denote the number of evaluation clips, K the number of classes, p^ik p_ik the predicted probability that clip i belongs to class k, and yiy_i the ground-truth label. We write y^i=argmaxkp^ik y_i= _k p_ik for the predicted label and c^i=maxkp^ik c_i= _k p_ik for its associated confidence. ECE measures the difference between confidence and accuracy across binned predictions. Partitioning the clips into M equal-width confidence bins B1,…,BMB_1,…,B_M over [0,1][0,1] according to c^i c_i, ECE is the sample-weighted mean absolute gap between per-bin accuracy and confidence: ECE=∑m=1M|Bm|n|acc(Bm)−conf(Bm)|,ECE= _m=1^M |B_m|n (B_m)-conf(B_m) , (13) where acc(Bm)=1|Bm|∑i∈Bm[y^i=yi],conf(Bm)=1|Bm|∑i∈Bmc^i.acc(B_m)= 1|B_m| _i∈ B_m1[ y_i=y_i], (B_m)= 1|B_m| _i∈ B_m c_i. (14) We use M=15M=15 bins and omit empty bins from the sum. A perfectly calibrated classifier satisfies acc(Bm)=conf(Bm)acc(B_m)=conf(B_m) for all m, yielding ECE=0ECE=0. The Brier score quantifies the squared difference between the predicted probability vector and the one-hot encoding of the true outcome: BS=1n∑i=1n∑k=1K(p^ik−[yi=k])2.BS= 1n _i=1^n _k=1^K ( p_ik-1[y_i=k] )^2. (15) Whereas ECE inspects only the top-label confidence, the Brier score is a strictly proper scoring rule that penalises the entire predictive distribution, jointly rewarding calibration and sharpness. Under this unnormalised, sum-over-classes convention BS∈[0,2]BS∈[0,2], with 0 attained when unit probability mass is placed on the correct class. Both quantities are computed per fold and averaged across the five folds. Calibration quality is critical for uncertainty-aware downstream applications such as selective prediction, human-in-the-loop verification, and risk-sensitive deployment scenarios. 6.1.5 Selective Prediction. We report selective-prediction operating points by sweeping the per-sample maximum-softmax confidence (separately for the country and the city head) and reporting the threshold and pooled metrics that achieve 80%80\% coverage of the validation set. 6.1.6 Mantel Test on the Latent Space. To quantify the relationship between linguistic and geographic distance, we run a permutation Mantel test on 1,124,2501,124,250 validation pairs (10,000 permutations), correlating Euclidean distance in poolh_pool with great-circle distance between ground-truth coordinates. 6.1.7 Border vs. Interior Decomposition. We additionally split the validation set into border and interior samples. In the absence of a country-polygon library on the training environment, we use a centroid-based heuristic: a sample is labelled border if any other country’s centroid is closer than 150150 km to the sample’s true coordinates. We report mean error, median error, and country accuracy for each subset, and treat the resulting numbers as an indicative rather than a definitive boundary analysis. 6.1.8 Unseen-City Zero-Shot Evaluation. The aim of this evaluation is to answer the single question that standard cross-validation cannot: how well does the model localize speech from a city it has never seen during training? Under an ordinary shuffled split, every city in the validation set is also present in the training set, so the model can succeed by memorizing city-specific acoustic signatures rather than learning the underlying geographic–dialectal structure. The resulting metrics therefore conflate genuine geographic generalization with city-level recall and systematically overstate deployable performance, since any real-world system will inevitably encounter speakers from locations absent in its training data. To isolate true generalization, for folds 1–4 we mask two cities per fold from the training set while retaining their samples in validation (Section 5.3.3), and evaluate these held-out cities separately from the seen cities of the same fold. We report the seen/unseen mean-error ratio as a direct measure of zero-shot geographic generalization: a ratio near 11 indicates that the model interpolates smoothly to unseen locations from neighbouring dialect regions — exactly the behaviour predicted by the continuum hypothesis — whereas a large ratio reveals reliance on having seen the target city. This protocol provides a leakage-free lower bound on real-world performance and, as Section 7.6 shows, exposes how strongly zero-shot accuracy depends on how densely the surrounding dialect region is sampled. 7 Results and Discussion All numbers reported in this section are pooled across the five validation folds described in Section 5.3.3; per-fold means and 95% bootstrap confidence intervals are given alongside the pooled value where appropriate. 7.1 Overall Geolocation Performance The proposed model attains a pooled mean localization error of 901.5901.5 km and a pooled median of 481.2481.2 km on the held-out validation samples (one prediction per source recording, aggregated across folds). The wide gap between mean and median, together with the 90%90\%-percentile error of 2340.52340.5 km, confirms the heavy-tailed nature of geographic prediction: the network localizes a clear majority of utterances to within a few hundred kilometres, but a long tail of confusable samples (low-content recordings, code-switching, or dialects with weak distinctive cues) drives the mean upward. The fold-level mean is 901.5±77.0901.5± 77.0 km (95% CI [854.6,969.1][854.6,969.1]) and the fold-level median is 472.8±32.9472.8± 32.9 km (95% CI [445.3,497.0][445.3,497.0]), indicating that fold-to-fold variability is small and the pooled numbers are stable. Figure 3 plots the full distribution of geodesic errors alongside these summary statistics, and Table 4 lists the headline metrics. Metric Value Unit Mean Error 901.5 km Median Error 481.2 km 25th percentile 179.9 km 75th percentile 1228.2 km 90th percentile 2340.5 km Mean (5-fold, 95% CI) [854.6,969.1][854.6,969.1] km Median (5-fold, 95% CI) [445.3,497.0][445.3,497.0] km Figure 3: Distribution of geodesic localization errors on the pooled validation set and corresponding summary statistics. The error distribution is strongly right-skewed: the median (481 km) sits well below the mean (902 km) and the 90th percentile (2341 km), reflecting a long tail of high-error samples. Threshold accuracies confirm that performance scales smoothly across geographic resolutions. As Figure 4 shows, 13.7%13.7\% of predictions fall within 100100 km, 51.7%51.7\% within 500500 km, 69.5%69.5\% within 10001000 km, and 90.9%90.9\% within 25002500 km. These numbers establish that the model reliably places most utterances within the correct multi-country region, while sub-city precision remains difficult under leakage-free evaluation. The auxiliary heads attain pooled top-1 accuracies of 64.5%64.5\% for country (5-fold mean 0.645±0.0440.645± 0.044, 95% CI [0.608,0.677][0.608,0.677]) and 45.2%45.2\% for city (0.452±0.0500.452± 0.050, CI [0.415,0.489][0.415,0.489]). Table 4: Headline geolocation performance, pooled across 5-fold GroupKFold validation. Metric Value Unit Pooled mean error 901.5 km Pooled median error 481.2 km Country accuracy (pooled) 64.5 % City accuracy (pooled) 45.2 % Threshold Accuracy Incremental Gain <50<50 km 4.8% 4.8% <100<100 km 13.7% 8.9% <250<250 km 32.2% 18.5% <500<500 km 51.7% 19.5% <1000<1000 km 69.5% 17.8% <2500<2500 km 90.9% 21.4% Figure 4: Cumulative distribution of localization errors and accuracy within increasing distance thresholds. 7.1.1 Per-Fold Decomposition To make the city-masking protocol reproducible, Table 5 reports per-fold validation size, masked cities, and pooled mean/median error. The five folds are remarkably consistent in median error (range 430430–507507 km, std only 3333 km), while the mean error is dominated by the heavy tail and varies more (range 840840–10331033 km). Folds with masked cities incur slightly larger mean errors than the unmasked baseline (fold 0) on average — a foretaste of the unseen-city analysis in Section 7.6. We also include in Figure 5 a side-by-side bar visualization of these per-fold means and medians against the pooled value, which makes the small fold-to-fold spread visible at a glance. Table 5: Per-fold validation size, cities masked from training, pooled mean/median geodesic error, and auxiliary head accuracies. The 5-fold means ± std are 901.5±77.0901.5± 77.0 km (mean), 472.8±32.9472.8± 32.9 km (median), 0.645±0.0440.645± 0.044 (country), and 0.452±0.0500.452± 0.050 (city). Fold Masked cities nvaln_val Mean Median Country City (km) (km) acc. acc. 0 – 466 901.1 490.3 0.691 0.474 1 Batna, El Obeid 466 1033.3 507.0 0.663 0.412 2 Makkah, Tripoli 466 860.0 446.1 0.618 0.461 3 Tunis, Sharjah 466 840.1 490.5 0.584 0.395 4 Cheikh Taba, Aleppo 465 872.7 430.0 0.671 0.518 Pooled – 2329 901.5 481.2 0.645 0.452 Figure 5: Per-fold mean and median geodesic error. Dashed lines indicate the pooled mean (901.5901.5 km) and median (481.2481.2 km). The remarkable stability of the median across folds (±33± 33 km) contrasts with the heavier-tailed mean (±77± 77 km) and reflects the heavy-tailed nature of geodesic error. 7.2 Country and City-Level Analysis 7.2.1 Country-Level Performance Per-country performance varies by an order of magnitude (Table 6). The model performs best on countries with dense, geographically compact city distributions, including Egypt (mean 317.2317.2 km, median 153.6153.6 km), Sudan (mean 326.9326.9 km, median 232.2232.2 km) and Algeria (mean 639.2639.2 km, median 307.1307.1 km despite its 507507-sample size and large within-country dispersion). Maghrebi and Gulf states with broader internal dialectal variation, such as Morocco (mean 1399.31399.3 km) and the United Arab Emirates (mean 1435.21435.2 km), show clearly weaker localization. Yemen (mean 2293.32293.3 km, median 1925.81925.8 km) is the worst-performing top-10 country, with most samples mislocated outside its borders entirely. Figure 6 visualizes these per-country mean and median errors as horizontal bars sorted by median, making the broad three-tier picture (Egypt/Sudan; Algeria/Jordan/Palestine/Kuwait/Saudi; Morocco/UAE/Yemen) immediately apparent. Table 6: Country-level geolocation performance (top 10 by validation samples). Acc@500 reports the fraction of samples localized within 500500 km of ground truth. Country Samples Mean (km) Median (km) Acc@500 Algeria 507 639.2 307.1 0.69 Sudan 290 326.9 232.2 0.82 Egypt 250 317.2 153.6 0.83 Morocco 229 1399.3 852.2 0.32 UAE 212 1435.2 1272.5 0.24 Saudi Arabia 144 893.3 633.7 0.39 Jordan 103 725.4 547.6 0.46 Palestine 100 768.0 432.4 0.52 Kuwait 81 850.3 596.2 0.44 Yemen 72 2293.3 1925.8 0.15 Figure 6: Per-country geodesic error (top 10 by validation samples), sorted by median. Sample-count annotations make the strong dependence on sample density explicit: the densely sampled Egypt/Sudan/Algeria triplet dominates the leaderboard, while sparser Maghrebi and Gulf countries trail by a wide margin. 7.2.2 City-Level Performance City-level results echo the country picture (Table 7). The strongest cities are Alexandria (182.5182.5 km mean, 97.997.9 km median — the only city with sub-100 km median), El Obeid in Sudan (291.8291.8 km mean, 203.3203.3 km median), Omdurman (332.9332.9 km), and Cairo (404.0404.0 km), all densely sampled and dialectally distinctive. Conversely, Fes (1285.81285.8 km), Ajman (1250.11250.1 km) and Sharjah (1373.71373.7 km) sit at the opposite end: Fes is consistently confused with other Moroccan cities, while the two UAE cities are dragged toward each other and toward Saudi alternatives. Algiers’ relatively large mean (792.5792.5 km) coexists with a much lower median (338.4338.4 km), an unmistakable signature of a small number of long-distance failure modes inflating the average. Figure 7 visualizes these city-level errors on the map, connecting each ground-truth location to its predicted counterpart so that the regional pattern of confusions is directly visible. Table 7: City-level geolocation performance (top 10 by validation samples). Acc@500 reports the fraction of samples localized within 500500 km of ground truth. City Samples Mean (km) Median (km) Acc@500 El Obeid 171 291.8 203.3 0.85 Cairo 152 404.0 198.4 0.76 alger 121 792.5 338.4 0.62 amman 103 725.4 547.6 0.46 Fes 101 1285.8 850.1 0.34 Alexandria 98 182.5 97.9 0.93 kuwait 81 850.3 596.2 0.44 Omdurman 80 332.9 227.9 0.90 Annaba 80 775.8 328.1 0.63 Ajman 73 1250.1 817.5 0.41 Figure 7: Geographic visualization of city-level audio geolocation performance on the pooled validation set. Blue markers indicate ground-truth city locations and red markers denote model-predicted locations; line segments connect each pair, making per-region error patterns directly visible. 7.3 Dialect Continuum Analysis 7.3.1 Mantel Test Results We performed a permutation Mantel test correlating Euclidean distance in poolh_pool with great-circle distance between ground-truth coordinates, evaluated on 1,124,2501,124,250 validation pairs (drawn from 1,5001,500 samples) and 10,00010,000 permutations. The test yields r=0.131r=0.131 with p≈1×10−4p≈ 1× 10^-4. The correlation is weak in absolute terms but extremely robust statistically, providing quantitative support for the dialect continuum hypothesis: linguistically similar utterances are systematically more likely to originate from geographically proximate locations. The modest magnitude is consistent with sociolinguistic expectations — non-geographic factors such as media exposure, education, migration, and individual variation contribute substantial linguistic distance that is decoupled from physical distance. Figure 8 plots learned linguistic distance against geographic distance for the sampled pairs, illustrating the weak but systematic positive association captured by the test. Figure 8: Relationship between geographic and learned linguistic distance for sampled validation pairs. The Mantel correlation r=0.131r=0.131 (p≈10−4p≈ 10^-4, 10,00010,000 permutations over 1,124,2501,124,250 pairs) is small but highly significant, evidence that geography is one — but not the dominant — structuring axis of the latent space. 7.3.2 Latent Space Structure A two-dimensional t-SNE / UMAP projection of poolh_pool coloured by latitude and longitude (Figure 9) shows smooth colour gradients rather than discrete clusters separated by sharp boundaries, qualitatively consistent with the Mantel finding. We deliberately do not report numerical “gradient smoothness” or “sharp boundary” statistics here because we no longer have a calibrated baseline against which their absolute values are meaningful; we treat the projection as a visualization of the structural claim made above rather than as standalone evidence. Figure 9: Two-dimensional projection of the learned latent space coloured by ground-truth latitude and longitude. The smooth colour gradients indicate that geographic information is encoded continuously rather than via discrete cluster boundaries. 7.4 Border vs. Interior Performance Table 8 reports performance on the centroid-based border / interior split described in Section 6.1.7. Under this heuristic, 360360 samples are flagged as border-adjacent and 1,9691,969 as interior. Border samples incur a modest but consistent 9%9\% higher mean localization error (967.7967.7 km vs. 889.4889.4 km, ratio 1.09×1.09×); the median gap is larger in absolute terms (688.3688.3 km vs. 435.6435.6 km), reflecting that border-adjacent samples are systematically harder to place precisely. The story is markedly more pronounced at the country head: country accuracy drops from 0.7020.702 in the interior to 0.3330.333 at the border — a 3737-point gap, more than a halving. In other words, the regression head degrades only modestly across boundaries (consistent with the continuum hypothesis), while the discrete country head suffers exactly the kind of catastrophic degradation one would predict for hard-boundary classifiers operating near isoglosses. Figure 10 contrasts the two error distributions, showing the heavy overlap in geodesic error against the sharp divergence in country accuracy. Table 8: Border vs. interior performance using a 150150 km nearest-other-country-centroid heuristic. The regression error is ∼9% 9\% higher at borders, but country-classifier accuracy degrades sharply (−37-37 percentage points). Subset Samples Mean (km) Median (km) Country Acc. Border 360 967.7 688.3 0.333 Interior 1969 889.4 435.6 0.702 Figure 10: Per-sample geodesic error stratified by border / interior status under the centroid-based heuristic. Distributions overlap heavily for the regression error (border =1.09×=1.09× interior on the mean) but differ markedly for country accuracy (0.3330.333 vs. 0.7020.702). 7.5 Calibration and Uncertainty Analysis 7.5.1 Calibration Quality The two auxiliary heads exhibit moderate calibration error: the country head reaches ECE=0.106ECE=0.106 with Brier score 0.4970.497, and the city head ECE=0.139ECE=0.139 with Brier score 0.7160.716. The numbers are stable across folds, suggesting the miscalibration is a property of the model class rather than of any single fold. Both heads tend toward over-confidence, a well-known consequence of cross-entropy training with label smoothing in low-coverage regimes; temperature scaling and isotonic regression are obvious post-hoc remedies that we leave to deployment-time work. The reliability diagrams in Figure 11 make this over-confidence visible, with both curves sitting below the diagonal across most confidence bins. Figure 11: Reliability diagrams for the country and city classification heads. ECE values of 0.1060.106 (country) and 0.1390.139 (city), with Brier scores of 0.4970.497 and 0.7160.716 respectively, indicate moderate over-confidence in both heads. 7.5.2 Selective Prediction Capability Even without explicit calibration, max-softmax confidence is a usable abstention signal. Table 9 reports the operating points obtained by sweeping the per-sample maximum-softmax confidence to retain 80%80\% of validation samples. Conditioning on country-head confidence at threshold τcountry=0.481 _country=0.481 raises country accuracy from 64.5%64.5\% to 75.3%75.3\% and pulls the pooled mean geodesic error from 901.5901.5 km down to 749.2749.2 km (median: 481.2→350.5481.2→ 350.5 km) — a 17%17\% reduction in mean error and 27%27\% reduction in median error for an acceptable abstention rate. Conditioning on city-head confidence (τcity=0.301 _city=0.301) yields a similar pattern at slightly weaker localization (mean 767.3767.3 km, median 365.6365.6 km) but pushes city accuracy on the kept subset above 53%53\%. Table 9: Selective prediction at 80%80\% coverage (nkept=1863n_kept=1863 out of 23292329). Threshold is chosen on the corresponding head’s max-softmax confidence; metrics are pooled across folds. Conf. head Threshold Country Acc. City Acc. Mean (km) Median (km) Country 0.481 0.753 0.527 749.2 350.5 City 0.301 0.724 0.534 767.3 365.6 7.6 Unseen-City Zero-Shot Evaluation The city-masking protocol (Section 5.3.3) lets us measure how the model behaves on cities that were never visible during training. Aggregating over the eight masked cities across folds 1–4, mean error on unseen-city samples is 1173.31173.3 km (n=99n=99) compared to 889.4889.4 km on seen-city samples of the same folds (n=2230n=2230). The ratio of 1.32×1.32× quantifies the cost of a strict zero-shot evaluation: substantial, but smaller than what one might expect for a system that has no in-distribution view of those cities. Per-city numbers (Table 10) reveal sharp heterogeneity. Some held-out cities collapse: Tripoli (2192.82192.8 km mean) is mislocated outside Libya entirely, Aleppo (1506.81506.8 km) is dragged toward the Maghreb, and batna’s mean (1463.31463.3 km) is dominated by a long tail even though its median (699.8699.8 km) shows that most predictions stay within Algeria. Sharjah (1345.21345.2 km) is consistently confused with other Gulf locations. Others degrade only mildly: tunis (497.7497.7 km mean) and El Obeid (597.0597.0 km) are essentially as accurate as a typical seen city, presumably because other cities of the same dialect region (alger, Annaba and the Sudanese cities Cairo / Omdurman, respectively) are present in training and span the relevant dialect neighborhood. This heterogeneity is the most informative observation in the paper: zero-shot performance is dominated by how well the rest of the training set covers the geographic and dialectal neighborhood of the held-out city, not by any intrinsic difficulty of that city. Table 10: Per-city zero-shot evaluation: each city is held out of training in exactly one fold and evaluated on its (still-present) validation samples in that fold. The right-most pair of columns reports the corresponding seen-city baseline from fold 0 (where available), where the same city was not masked. Held-out city Fold nunseenn_unseen Mean (km) Median (km) Seen mean (km, fold 0) nseen, f0n_seen, f0 Tripoli 2 12 2192.8 2340.8 1581.6 10 Aleppo 4 6 1506.8 1303.2 1169.5 6 Batna 1 12 1463.3 699.8 233.9 9 Sharjah 3 29 1345.2 1365.9 – 0 Makkah 2 1 1150.1 1150.1 1099.3 8 Cheikh Taba 4 4 669.9 713.4 759.2 2 El Obeid 1 30 597.0 485.3 237.5 44 Tunis 3 5 497.7 482.7 487.3 1 Pooled (unseen, n=99n=99) – 99 1173.3 – – – Pooled (seen, n=2230n=2230) – 2230 889.4 – – – Ratio – – 1.32×1.32× – – – 7.7 Discussion and Implications 7.7.1 Theoretical Implications Our results provide qualitative support for the Arabic dialect continuum hypothesis, but with important quantitative caveats. The Mantel test on the latent space is small in magnitude (r=0.131r=0.131) yet highly significant; t-SNE projections show smooth geographic gradients rather than discrete clusters; and the regression head’s mean error increases by only 9%9\% for border-adjacent samples relative to interior ones, even though the discrete country head shows the large 3737-point border degradation expected of any hard-boundary classifier. Taken together these observations are consistent with a continuum, but they do not by themselves rule out a discrete-but-noisy generative process; the present work establishes the leakage-free baseline against which future, more controlled tests of the continuum hypothesis can be run. 7.7.2 Practical Applications Beyond theoretical insights, our approach enables several practical applications. First, selective prediction allows deployment in scenarios requiring high reliability, with the system abstaining when confidence is insufficient. Second, hierarchical predictions (country, city, coordinates) provide graceful degradation and interpretable outputs across multiple granularities. Third, the continuous nature of predictions enables applications requiring precise geographic resolution, such as forensic speaker analysis or targeted content localization. The model’s learned representations could also support downstream tasks. The latent space poolh_pool encodes dialectal information that may improve speech recognition, machine translation, or dialect adaptation systems. Transfer learning from our pre-trained encoder could accelerate development of specialized dialect processing tools. 8 Limitations This work is subject to several limitations related to data quality, modeling assumptions, and evaluation methodology. Data and Annotation Limitations. The dataset exhibits strong geographic imbalance, with many cities represented by very few samples and rural dialects largely absent; underrepresented countries (Qatar, Bahrain, Oman) and rural dialects consequently receive much less accurate predictions. This limits generalization to underrepresented regions. Additionally, geographic labels are approximated using city centroids derived from self-reported origins, introducing unavoidable localization error and preventing fine-grained spatial accuracy. Migration histories and dialectal accommodation are not explicitly modeled, further affecting label reliability. Speakers may also be influenced by more than one dialect, for instance expatriates or individuals of mixed heritage, and such multi-dialect influence is not captured by the single-origin labeling scheme. The border/interior split used in our analysis (Section 7.4) further relies on a centroid-distance heuristic rather than true country polygons, adding an additional source of label imprecision. Model Limitations. The model produces point coordinate estimates without uncertainty quantification, and auxiliary classifiers are poorly calibrated, limiting reliability in downstream or high-stakes use cases. Calibration quality, while reasonable (ECE ≈0.11≈ 0.11 country, ≈0.14≈ 0.14 city), is not production-ready and would benefit from temperature scaling or isotonic regression at deployment. Moreover, despite hierarchical outputs, the learned acoustic representations remain largely uninterpretable, restricting linguistic insight beyond aggregate performance. Performance on cities unseen during training (Section 7.6) also degrades substantially, by 1.32×1.32× on average, with very high variance: cities surrounded by other in-vocabulary cities of the same dialect region (e.g., Tunis, El Obeid) degrade only mildly, while geographically isolated cities (Tripoli, Aleppo, Sharjah in our protocol) degrade severely. Addressing this would require either denser coverage in underrepresented regions or explicit modelling of dialect-region prototypes. Linguistic Limitations. Code-switching between dialects and Modern Standard Arabic weakens dialect-specific cues, reducing localization accuracy; more generally, the model relies on dialectal features that are attenuated by code-switching and formal register, making such utterances systematically harder to localize. Within-city variation due to socioeconomic, generational, and stylistic factors is not captured, as annotations assume dialectal homogeneity at the city level. Evaluation Limitations. Geodesic distance metrics do not account for dialectal similarity, treating all spatial errors equally regardless of linguistic relevance. Furthermore, the absence of standardized benchmarks for Arabic dialect geolocation limits direct comparison with prior work. 9 Conclusion We have presented a regression-based approach to Arabic dialect geolocation that models dialectal variation as a continuous geographic space rather than discrete categories. By formulating the task as regression over spherical coordinates and combining frame-level XLS-R-300M and Whisper-large-v3 features with phonotactic descriptors through a Transformer encoder and a learnable attention-pool query, our system avoids the classification–quantization error of city-bin baselines and the planar distortion of latitude–longitude regression. Under a leakage-free 5-fold GroupKFold protocol grouped by source recording, our model attains a pooled mean error of 901.5901.5 km and a median error of 481.2481.2 km, with 51.7%51.7\% of predictions within 500500 km, 69.5%69.5\% within 10001000 km, and 90.9%90.9\% within 25002500 km of ground truth. Auxiliary country and city heads reach pooled accuracies of 64.5%64.5\% and 45.2%45.2\%, respectively, and a confidence-driven selective-prediction policy at 80%80\% coverage further reduces mean error from 901.5901.5 km to 749.2749.2 km (median: 481.2→350.5481.2→ 350.5 km). A permutation Mantel test on the learned latent space (r=0.131r=0.131, p≈10−4p≈ 10^-4) provides quantitative, leakage-free support for the Arabic dialect continuum hypothesis. Methodologically, one of the most important contributions of this paper is the city-masking cross-validation protocol, which lets us measure true zero-shot generalization to held-out cities. Mean error on unseen cities rises to 1173.31173.3 km (a 1.32×1.32× degradation relative to seen cities) and the spread across cities is dominated by how well the rest of the training set covers the held-out city’s dialect region. This is, to our knowledge, the first leakage-free zero-shot evaluation of Arabic dialect geolocation on this corpus. Future work should focus on direct comparisons against discrete classification baselines under the same leakage-free protocol, polygon-based border analyses, expanded geographic coverage of underrepresented dialects, and post-hoc confidence calibration for deployment. Acknowledgements.The authors would like to thank Prince Sultan University for their support. References A. Abdelali, H. Mubarak, et al. (2021) QADI: Arabic dialect identification in the wild. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, p. 1–10. Cited by: §2.5. M. Abdul-Mageed, A. Elmadany, et al. (2021) ARBERT & marbert: deep bidirectional transformers for arabic. In Proceedings of the 59th annual meeting of the association for computational linguistics and the 11th international joint conference on natural language processing (volume 1: long papers), p. 7088–7105. Cited by: §2.4. M. Abdul-Mageed, A. Elmadany, et al. (2024) NADI 2024: the fifth nuanced Arabic dialect identification shared task. In Proceedings of ArabicNLP 2024, Cited by: §2.4, §2.5, Table 1. A. Ali, N. Dehak, P. Cardinal, S. Khurana, S. H. Yella, J. Glass, P. Bell, and S. Renals (2016) Automatic dialect detection in Arabic broadcast speech. In Proc. Interspeech 2016, p. 2934–2938. Cited by: §2.2, Table 1. W. Antoun, F. Baly, and H. Hajj (2020) AraBERT: transformer-based model for Arabic language understanding. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, p. 9–15. Cited by: §2.4. A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y. Saraf, J. Pino, A. Baevski, A. Conneau, and M. Auli (2022) XLS-R: self-supervised cross-lingual speech representation learning at scale. In Proc. Interspeech 2022, p. 2278–2282. Cited by: §2.2, §2.7, Figure 2. H. Bouamor, N. Habash, M. Salameh, W. Zaghouani, O. Rambow, et al. (2018) The MADAR Arabic dialect corpus and lexicon. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC), p. 3387–3396. Cited by: §2.1, §2.5, Table 1. J. K. Chambers and P. Trudgill (1998) Dialectology. 2 edition, Cambridge Textbooks in Linguistics, Cambridge University Press. Cited by: §1. M. Elaraby and M. Abdul-Mageed (2018) Deep models for Arabic dialect identification on benchmarked data. In Proceedings of the Fifth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial), p. 263–274. Cited by: §2.4. H. Elfardy and M. Diab (2013) Sentence level dialect identification in Arabic. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, p. 456–461. Cited by: §2.1. N. Habash, M. T. Diab, and O. Rambow (2012) Conventional orthography for dialectal Arabic. In Proceedings of LREC, p. 711–718. Cited by: §2.6. I. Hamed et al. (2025) A survey of code-switched arabic nlp. arXiv preprint arXiv:2501.13419. Cited by: §2.6. W. Heeringa and J. Nerbonne (2001) Dialect areas and dialect continua. Language Variation and Change 13 (3), p. 375–400. Cited by: §1. A. Keleg, S. Goldwater, and W. Magdy (2023) ALDi: quantifying the Arabic level of dialectness of text. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, p. 10597–10611. Cited by: §2.3, Table 1. S. Khurana, M. Najafian, A. Ali, T. Al Hanai, Y. Belinkov, and J. Glass (2017) Multi-view dimensionality reduction for dialect identification of Arabic broadcast speech. In Proc. Interspeech 2017, p. 2596–2600. Cited by: §2.2. O. Nacar, S. Sibaee, A. Ammar, Y. Alhabashi, N. S. Sibai, Y. F. Ahmed, A. S. Alqusaiyer, S. M. AlMahmoud, A. M. Mukhaniq, L. Raed, S. M. Alatwah, W. N. Alqahtani, Y. A. Alnasser, M. A. Khadraoui, and W. Boulila (2026) ARCADE: a city-scale corpus for fine-grained Arabic dialect tagging. arXiv preprint arXiv:2601.02209. Cited by: 3rd item, §2.2, Table 1, §4.1, §4.2, §4.4. J. Nerbonne (2009) Data-driven dialectology. Language and Linguistics Compass 3 (1), p. 175–198. Cited by: §1. O. Obeid et al. (2020) CAMeL tools. In Proceedings of LREC, p. 7022–7032. Cited by: §2.7. OpenStreetMap Foundation (2024) Nominatim: open-source geocoding with OpenStreetMap data. Note: https://nominatim.org/Accessed 21/07/2026 Cited by: §4.1, §4.4. A. Pasha et al. (2014) MADAMIRA. In Proceedings of LREC, p. 1094–1101. Cited by: §2.7. A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever (2023) Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning (ICML), p. 28492–28518. Cited by: §2.2, §2.7, Figure 2. S. Sha’ban and N. Habash (2025) The arabic generality score: another dimension of modeling arabic dialectness. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, p. 29990–30001. Cited by: §2.3. S. Shon, A. Ali, Y. Samih, H. Mubarak, and J. Glass (2020) ADI17: a fine-grained Arabic dialect identification dataset. In Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 8244–8248. Cited by: §2.2, Table 1. A. B. Soliman et al. (2017) AraVec. In Procedia Computer Science, Vol. 117, p. 256–265. Cited by: §2.7. B. Talafha et al. (2024) Casablanca. arXiv preprint arXiv:2410.04527. Cited by: §2.2. V. Vivanco Cepeda, G. K. Nayak, and M. Shah (2023) GeoCLIP: clip-inspired alignment between locations and images for effective worldwide geo-localization. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 36, p. 8690–8701. Cited by: Figure 2. O. F. Zaidan and C. Callison-Burch (2011) The Arabic online commentary dataset. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, p. 37–41. Cited by: §2.5. O. F. Zaidan and C. Callison-Burch (2014) Arabic dialect identification. Computational Linguistics 40 (1), p. 171–202. Cited by: §2.1, Table 1.