1. Introduction
Digital language technologies have transformed how we interact with information, yet their development remains highly uneven across the world’s languages. While resource-rich languages benefit from sophisticated processing tools, many historically marginalized languages face a growing “digital language divide” [1]. Bridging this divide is not merely a technical challenge but a crucial step in preserving linguistic diversity and ensuring equitable access to the digital ecosystem.
The Amazigh language (also known as Berber) represents a compelling case study in this context. As an indigenous language family of North Africa with ancient roots, Amazigh has substantial cultural and historical significance, with speaker communities spanning Morocco, Algeria, Libya, Tunisia, and several Sahelian countries. Recent estimates suggest between 30 and 40 million speakers across this region [2], making it one of the most widely spoken indigenous language families in Africa. Despite this substantial speaker base, Amazigh has historically been marginalized in both official policy and technological development.
The past two decades have witnessed significant changes in the sociopolitical status of Amazigh languages. Morocco’s 2011 constitutional recognition of Amazigh as an official language alongside Arabic marked a pivotal shift, followed by similar recognition in Algeria in 2016. These developments have catalyzed renewed interest in Amazigh language preservation, education, and digital integration. This period of increased digital engagement, described by researchers as a “digital awakening”, has created both opportunities and challenges for technological development [3]. The development of comprehensive language technologies for Amazigh presents several distinct challenges. Linguistically, Amazigh features rich morphological systems with templatic patterns and extensive affixation. This complexity exceeds that of many well-resourced languages, necessitating specialized computational approaches [4]. The language family’s dialectal diversity introduces further complications, with numerous varieties exhibiting significant phonological, lexical, and grammatical differences. This diversity challenges the development of unified processing systems and requires careful consideration of cross-dialectal applicability.
Amazigh’s orthographic variation presents additional challenges, as the language employs three distinct writing systems (Tifinagh, Latin-based, and Arabic-based), each with its own computational processing requirements. This orthographic diversity necessitates specialized approaches for text processing and recognition. These technical challenges are compounded by persistent resource scarcity, as Amazigh remains a low-resource language in computational terms, with limited availability of large-scale corpora, parallel texts, and standardized processing tools.
Developing effective language technologies requires considering not only technical implementation but also community adoption, educational applications, and sociolinguistic factors. Research indicates that these integrated factors play a crucial role in determining technological success and impact [5]. For Amazigh language processing, these considerations have influenced a distinctive research evolution. Initial efforts focused on creating fundamental linguistic resources, which subsequently enabled the development of advanced computational methods. This trajectory parallels methodological advances in computational linguistics while maintaining focus on Amazigh-specific challenges. The current survey examines this technological progression through systematic analysis of four core areas: Natural Language Processing (NLP), Speech Technologies, Optical Character Recognition (OCR), and Machine Translation. Previous research has explored individual components of Amazigh computational processing [6,7,8,9]. However, this study provides the first integrated examination across multiple domains and methodological frameworks.
Our contributions include a systematic analysis of methodological developments in Amazigh language technology from 2010 to 2025, tracing the evolution from rule-based to neural approaches across multiple domains. We examine the domain-specific challenges and innovations in morphological analysis, part-of-speech tagging, named entity recognition, speech recognition, optical character recognition, and machine translation. The survey assesses resource development initiatives and their impact on technological capabilities, analyzes dialectal coverage and script support across different domains, and identifies persistent challenges and promising research directions for future development.
Our analysis reveals both significant achievements and remaining gaps in Amazigh language technology. Although certain domains such as morphological analysis and optical character recognition have reached high levels of maturity, others like continuous speech recognition and machine translation remain in earlier developmental stages. This uneven development reflects both technological challenges and resource allocation patterns.
The remainder of this survey is organized as follows. Section 2 provides background on the Amazigh language and our methodology. Section 3 examines natural language processing tasks, including morphological analysis, part-of-speech tagging, and named entity recognition. Section 4 analyzes speech technologies, focusing on feature extraction, acoustic modeling, and recent neural approaches. Section 5 covers optical character recognition for both printed and handwritten Tifinagh text. Section 6 addresses the development of machine translation. Section 7 surveys available resources and datasets, while Section 8 examines practical applications and societal impact. Section 9 concludes with a synthesis of findings and future research directions.
2. Background and Methodology
2.1. The Amazigh Language: Characteristics and Computational Challenges
The Amazigh language (also known as Berber) belongs to the Afroasiatic language family and is indigenous to North Africa, with speaker communities across Morocco, Algeria, Libya, Tunisia, and several Sahelian regions. Following decades, exemplified by Morocco’s 2011 constitutional establishment of Amazigh as an official language alongside Arabic.
Figure 1 illustrates the geographic distribution of major Amazigh varieties across North Africa, highlighting the extensive territorial coverage and dialectal diversity that creates computational challenges for unified language processing systems. Amazigh comprises several major varieties, including Tarifit (northern Morocco), Tachelhit (southern Morocco), Central Atlas Tamazight (central Morocco), and Kabyle (northern Algeria). This dialectal diversity creates significant challenges for computational processing, particularly for cross-dialectal applications. The Ethnologue identifies thirteen distinct Amazigh languages, though the boundaries between varieties remain subject to ongoing linguistic debate.
A defining characteristic of Amazigh is its orthographic diversity, with three distinct writing systems in active use: Tifinagh: The indigenous script officially adopted by Morocco’s Royal Institute of Amazigh Culture (IRCAM) in 2003, shown in Figure 2. Latin-based: Various Latinized transcription systems used primarily in academic contexts and in Kabylia. Arabic-based: Adaptations of the Arabic script used in traditional and religious contexts.
From a computational perspective, Amazigh presents several distinctive challenges:
Morphological complexity: Amazigh employs templatic morphology with extensive affixation. The verbal system features complex inflectional paradigms encoding person, number, gender, aspect, mood, and negation. This complexity necessitates specialized processing approaches beyond those developed for Indo-European languages.
Syntactic structure: Amazigh typically follows VSO (verb-subject-object) word order, though with considerable dialectal variation. This differs from the SVO structure common in major resource-rich languages, creating challenges for computational parsing and generation.
Phonological variation: Amazigh dialects exhibit significant phonological differences, with phoneme inventories varying across regions. These variations manifest in both spoken language technologies and orthographic representation.
Resource scarcity: Despite recent advances in resource development, Amazigh remains under-resourced compared to major world languages, with limited availability of large-scale corpora, comprehensive lexicons, and standardized processing tools.
Figure 3 illustrates how these challenges have been addressed through technological evolution from 2010 to 2025, showing the progression from rule-based to statistical and neural approaches across four key domains.
2.2. Current State of Technology Development
Table 1 presents the current state of Amazigh language technology development across eight key domains. The assessment considers technological maturity, resource availability, empirical performance, dialectal coverage, and research activity.
As the table illustrates, significant disparities exist in developmental status across domains. Morphological analysis, POS tagging, and printed Tifinagh OCR demonstrate high maturity levels with strong performance metrics. In contrast, continuous speech recognition and machine translation remain at earlier developmental stages with more limited resources and dialectal coverage.
The resource landscape has been shaped by both institutional and academic contributions. The Royal Institute of Amazigh Culture has led standardization efforts for Tifinagh and provided foundational lexical resources. Academic research has focused on developing specialized corpora and task-specific datasets, though often limited to specific dialects or domains.
2.3. Survey Methodology
This survey systematically examines technological advances in Amazigh language processing across four primary domains: Natural Language Processing, Speech Technologies, Optical Character Recognition, and Machine Translation. Our analysis traces the methodological evolution from rule-based to neural approaches while addressing the unique computational challenges, assessing resource development initiatives, and identifying promising research directions.
We used a comprehensive literature selection strategy that focused on peer-reviewed publications from 2010 to 2025. Our approach combined database queries across major digital libraries, using terms combining “Amazigh”, “Berber”, and “Tamazight” with technology-specific keywords; backward and forward citation analysis from seminal papers and manual review of proceedings from conferences on under-resourced languages. The selection criteria prioritized direct relevance to Amazigh processing, methodological innovation, empirical validation, and contribution to resource development, with particular attention to research addressing dialectal variation, morphological complexity, and script diversity.
Our multidimensional evaluation framework examined the following: Technological progression: Tracing methodological transitions across three phases: rule-based approaches (2010–2015), statistical models (2015–2020), and neural architectures (2020–2025). Performance evaluation: Analyzing reported empirical results and benchmarks. Resource development: Assessing dataset creation, annotation efforts, and standardization initiatives. Dialectal coverage: Examining support for different Amazigh varieties. Application domains: Evaluating real-world implementations and practical applications.
This framework enabled a systematic comparison across technologies, the identification of common patterns, and the synthesis of insights on the unique trajectory of Amazigh language technology development. The temporal scope captures the complete evolution, from early exploratory efforts to contemporary neural approaches, providing a perspective to identify trends and project future directions.
2.4. Dialectal Coverage and Resource Allocation
The dialectal diversity of Amazigh creates significant challenges for equitable technology development, as linguistic differences across varieties require either separate models or sophisticated cross-dialectal approaches. Current resource allocation patterns favor institutionally supported varieties, creating performance disparities that may exacerbate digital divides within Amazigh communities. Table 2 illustrates the current resource distribution across major varieties.
Current allocation heavily favors Central Atlas Tamazight (60% of resources), creating performance disparities. Systems show 15–30% accuracy degradation for under-resourced dialects [10].
Balancing Strategies: Cross-dialectal transfer learning [11], multi-task learning frameworks, and community-driven resource development can address these imbalances. A tiered allocation approach (40% unified models, 35% major dialects, and 25% underserved varieties) could improve equity while maintaining development efficiency.
2.5. Research Distribution and Publication Trends
The analysis of Amazigh language technology publications (2010–2025) reveals distinct patterns across domains and time periods. Table 3 illustrates the distribution of research efforts across the four primary technological domains.
The distribution reveals that NLP tasks and OCR dominate the research landscape (60% combined), while machine translation remains significantly underexplored (7%). The temporal analysis shows three distinct phases: foundation building (2010–2015, 25% of publications), adoption of statistical methods (2016–2020, 40% of publications) and integration of neural approaches (2021–2025, 35% of publications).
3. Natural Language Processing Tasks
The last decade has witnessed remarkable progress in Natural Language Processing for Amazigh. Researchers have gradually tackled the unique computational challenges posed by this historically significant North African language, moving from basic rule-based approaches to sophisticated neural architectures that rival those used for resource-rich languages.
3.1. Core NLP Technologies: Evolution and Current State
3.1.1. Morphological Analysis
Amazigh’s complex morphological system presents substantial computational challenges. With its templatic patterns, extensive affixation, and rich inflectional paradigms, processing Amazigh morphology requires specialized approaches beyond those developed for Indo-European languages.
The foundations of computational morphology for Amazigh were established in 2012 when Raiss and Cavalli-Sforza introduced ANMorph, the first system specifically designed for Amazigh noun analysis [12]. This groundbreaking work demonstrated that rule-based approaches could effectively capture the complex patterns of Amazigh nominal morphology. Building on this foundation, Nejme et al. [13] developed a comprehensive finite-state morphology framework that expanded coverage to other word classes.
AmAMorph [14], a sophisticated finite-state analyzer, significantly improved coverage and accuracy. Around the same time, researchers began developing specialized systems for specific morphological phenomena. Notably, Ammari and Zenkoua created APMorph [15], which focused exclusively on pronominal morphology, an area of particular complexity in Amazigh.
Dialect-specific variations received attention through targeted studies, such as Oussou’s analysis of Tashlhiyt morphology in the Ayt Hdidou community [16]. This work highlighted the challenges of handling regional variations within a unified computational framework.
Recent years have seen a shift toward hybrid approaches that combine traditional linguistic knowledge with data-driven methods. Neural enhancements to morphological analyzers have shown promise in improving both coverage and accuracy, particularly for handling dialectal variations and morphological ambiguity. This integration represents a promising direction for addressing the persistent challenges in Amazigh morphological analysis.
3.1.2. Part-of-Speech Tagging
POS tagging for Amazigh represents the complete methodological evolution seen across NLP as a field. The journey began with fundamental linguistic work by Outahajala et al. [17], who developed a fine-grained tagset specifically designed for Amazigh’s unique morphosyntactic properties. This linguistic groundwork provided the essential framework for subsequent technological development.
The field then progressed to statistical machine learning approaches. Amri and Zenkouar made significant contributions in this area, introducing methods that could learn tagging patterns from annotated corpora [18]. Their language-independent model [19] marked a significant breakthrough, achieving competitive performance while reducing the need for extensive language-specific engineering. These advances were made possible by the creation of morphosyntactically annotated corpora [20], highlighting the critical relationship between resource development and technological advancement.
The neural revolution arrived in the 2020s, with several teams exploring different deep learning architectures. Maarouf pioneered the application of LSTM networks for processing Tifinaghe text [21], while GRU architectures were investigated for their computational efficiency [22]. The current state-of-the-art performance was established by Bani et al. in 2023, whose deep neural networks achieved impressive accuracy rates of 93.8% [23].
Comparative studies by Amri et al. [24] have validated the advantages of neural approaches over statistical methods, showing consistent improvements across different evaluation metrics. This progression from rule-based to statistical to neural approaches mirrors the broader evolution in NLP while addressing the specific challenges of Amazigh.
3.1.3. Named Entity Recognition
The development of NER systems for Amazigh tells a story of gradual sophistication, from simple rule-based frameworks to hybrid systems combining linguistic knowledge with statistical learning. Boulaknadel et al. pioneered this area in 2014 with the first rule-based framework [25], followed closely by Talha et al.’s NERAM system [26], which focused on identifying core entity types in Amazigh text.
As the field matured, hybrid approaches emerged that combined rule-based components with statistical methods. Talha et al. continued to make significant contributions, developing SVM-based solutions [27] and specialized hybrid frameworks that addressed the challenges of named entity disambiguation in context [28]. Their work on enhanced hybrid systems [29] demonstrated significant performance improvements in different types of entities.
In parallel, Amri et al. explored alternative hybrid statistical approaches [30], contributing to a growing body of knowledge on effective NER strategies for Amazigh. While progress has been substantial, the development of NER systems continues to be constrained by the limited availability of annotated data for training and evaluation. Innovative approaches to data augmentation and semi-supervised learning have helped researchers make progress despite these resource limitations.
3.1.4. Syntactic and Semantic Analysis
While morphological analysis and POS tagging have seen substantial development, syntactic and semantic processing remain at earlier stages of maturity. Nejme et al. [31] laid the groundwork in 2012, highlighting the challenges posed by Amazigh’s predominantly verb-initial word order and complex agreement patterns.
A significant leap forward came with Talha’s 2019 doctoral work on automated syntactic-semantic analysis for Amazigh [32]. Her system integrated morphological and syntactic processing while addressing semantic relationships in text, demonstrating effective handling of complex syntactic structures while maintaining semantic coherence. This work represented an important step toward a comprehensive linguistic analysis of Amazigh.
The relatively limited progress in this area reflects both the foundational nature of syntactic and semantic processing, which builds upon more basic NLP tasks, and the substantial challenges involved in modeling the complex interface between Amazigh syntax and semantics. This domain represents a critical frontier for future research. Table 4 summarizes the evolution of these Amazigh NLP technologies across three distinct developmental periods.
3.2. Challenges and Future Directions
Despite impressive progress, several key challenges continue to shape the direction of Amazigh NLP research:
Resource limitations remain a primary constraint, though recent work has made substantial headway in addressing this gap. Boulaknadel and Ataa Allah established crucial guidelines for standard corpus development [33], providing valuable frameworks for consistent resource creation. More recently, Maarouf’s comprehensive linguistic dataset [34] offers crucial support for multiple processing tasks, while Bani et al. created the first Amazigh lemmatizer [37]—another important addition to the resource landscape.
Dialectal variation presents persistent challenges that researchers continue to tackle. Faouzi et al. have shown promising results with word embeddings across dialectal varieties [38], while Oussou’s analysis of dialect-specific morphology [16] highlights the importance of accounting for regional differences. Ataa Allah and Boulaknadel have identified the development of robust cross-dialectal models as a key priority for future research [39].
Cross-lingual approaches have emerged as a clever strategy for addressing resource limitations. Diab et al. demonstrated effective results with guided back-translation between Kabyle and French [11], while Maarouf et al.’s work on transformer-based approaches to English-Amazigh translation [36] shows the potential of transfer learning from resource-rich languages.
Neural model limitations present additional challenges in low-resource contexts. While achieving state-of-the-art performance, neural approaches require substantially larger datasets than currently available: machine translation needs 1M+ parallel sentences versus 50K available, speech recognition requires 1000+ h vs. 100 h accessible. Computational constraints include limited GPU access and mobile deployment challenges. Mitigation strategies include transfer learning [36], data augmentation [11], and hybrid architectures [40], though substantial gaps remain.
Evaluation metric limitations present another significant challenge, as standard metrics may inadequately capture Amazigh’s linguistic complexity. BLEU scores for machine translation fail to account for morphological variants, dialectal differences, and flexible word order patterns, potentially penalizing valid translations. Similarly, WER for speech recognition does not distinguish between dialectal pronunciation variants and actual errors, while NLP classification metrics may not reflect cultural appropriateness or cross-dialectal applicability. Current research typically relies on single-reference evaluation despite multiple valid expressions across dialects. Future work should explore Amazigh-specific metrics incorporating morphological awareness, cultural context validation, and multi-dialectal reference sets to provide more meaningful assessment of system performance [11,36].
Looking at future directions, several paths appear particularly promising: Integration of traditional linguistic knowledge with modern neural methods. Development of sophisticated cross-dialectal models. Expansion of syntactic and semantic processing capabilities. Creation of standardized evaluation benchmarks. Application of self-supervised learning to maximize the utility of limited labeled data.
The convergence of traditional linguistic expertise with modern computational methods offers exciting possibilities for future development, as noted by Ataa Allah and Boulaknadel [39]. This synergy between linguistic knowledge and technological innovation will likely continue to drive advances in processing this historically significant language, ensuring its continued vitality in modern computing.
4. Speech Technologies
Speech technology research for the Amazigh language has demonstrated significant advancement in recent years, characterized by the methodological transition from traditional statistical approaches to contemporary neural architectures. This evolution reflects both broader trends in speech processing and specific challenges inherent to this historically under-resourced language with its distinct phonological characteristics and dialectal variations.
4.1. Development Trajectory and Methodological Advances
The initial phase of Amazigh speech recognition research focused primarily on establishing baseline systems and foundational resources. Satori and ElHaoussi [41] implemented the first systematic application of Hidden Markov Models (HMM) utilizing the CMU toolkit, while El Ouahabi et al. [42] developed an Amazigh-Tarifit Automatic Speech Recognition (ASR) system comprising 187 isolated words recorded by 50 speakers with equal gender distribution. The latter achieved a Word Error Rate (WER) of 8.20% using Gaussian Mixture Models (GMM) with Hidden Markov Models, thereby establishing initial performance benchmarks.
Telmem and Ghanou subsequently investigated parameter optimization for HMM-based approaches [43], whose systematic analysis of HMM configurations demonstrated the impact of architectural decisions on recognition performance. Concurrently, El Ouahabi et al. [44] addressed the critical need for linguistic resources through the development of AMZSRD (Amazigh Speech Recognition Database), a standardized corpus incorporating speaker diversity and dialectal variations.
The transition to neural methodologies commenced with comparative studies juxtaposing traditional and emerging approaches. Telmem and Ghanou [45] conducted a rigorous comparison between HMM and Convolutional Neural Network (CNN) acoustic models, demonstrating the superior representational capacity of the latter. This foundational work was extended by Telmem and Ghanou [46], whose CNN-based system achieved 93.9% accuracy for speakers aged 30 and above when evaluated on a substantial dataset comprising 9240 audio files for gender analysis and 13,860 files for age-based analysis.
Parallel investigations into specific application domains included Hamidi et al.’s [47] development of an interactive digit recognition system optimized for noisy environments, and Boulal et al.’s [48] CNN-based digit recognition system. The latter achieved 91.75% accuracy, 93% precision, and 92% recall when evaluated with 42 native speakers, demonstrating the efficacy of convolutional architectures even with constrained training data.
A significant methodological advancement was realized through the introduction of hybrid architectures integrating convolutional and recurrent components. Daouad et al. [40] conducted a comparative analysis of one-dimensional and two-dimensional CNN-LSTM architectures for isolated word recognition in the Tarifit dialect. Their investigation, utilizing a corpus comprising 2400 audio files from 80 speakers covering 30 isolated words, demonstrated that two-dimensional CNN-LSTM configurations yielded superior performance, achieving accuracy rates exceeding 96%.
This hybrid approach was further validated by Telmem et al. [10], whose systematic evaluation confirmed the efficacy of CNN-LSTM architectures across demographic categories. Their analysis revealed significant performance differentials across both gender and age cohorts, underscoring the importance of demographic representation in speech corpus development.
4.2. Feature Extraction and Acoustic Modeling
Feature extraction methodologies have constituted a critical dimension of Amazigh speech recognition research. Telmem et al. [49] conducted a comprehensive comparative analysis of MFCC, spectrogram, and Mel-spectrogram approaches. Their investigation revealed that Mel-spectrogram features consistently outperformed traditional MFCC features, particularly in capturing the distinctive phonological characteristics of Amazigh dialects.
The specific configuration of feature extraction parameters and their relationship to classifier performance was systematically investigated by Boulal et al. [48]. Their CNN-based system, utilizing optimized MFCC features, demonstrated the capacity of carefully engineered feature extraction to achieve high performance metrics even with limited training data.
Building upon these findings, Boulal et al. [50] introduced an enhanced feature extraction framework specifically calibrated for Amazigh speech patterns. Their methodology incorporated advanced preprocessing techniques addressing the unique challenges posed by Amazigh phonology, resulting in statistically significant improvements in recognition accuracy across varied speaking styles and dialectal variations.
4.3. Recent Innovations and Emerging Paradigms
Contemporary research in Amazigh speech recognition has exhibited three principal trajectories of innovation, each addressing specific limitations of prior approaches:
4.3.1. Advanced Neural Architectures
The integration of transformer architectures with convolutional networks has emerged as a promising direction for enhanced performance. Daouad et al. [51] developed a parallel CNN transformer-encoder model specifically designed for Amazigh speech recognition. This architectural configuration demonstrated significant performance improvements for continuous speech recognition, effectively combining the local feature extraction capabilities of CNNs with the long-range dependency modeling of transformers.
The optimization of connectionist temporal classification (CTC) approaches has been investigated by Telmem [52], whose CNN-CTC model achieved a favorable balance between computational efficiency and recognition accuracy. This architecture demonstrated particular efficacy in addressing the alignment challenges inherent to Amazigh speech recognition, especially for connected speech patterns.
4.3.2. Transfer Learning Approaches
The adaptation of large-scale pretrained models has been explored as a strategy for mitigating data scarcity. Daouad et al. [53] conducted a systematic investigation of Whisper model adaptation for Amazigh ASR, demonstrating competitive performance with significantly reduced training data requirements relative to traditional approaches. Their methodology emphasized the potential of cross-lingual transfer in low-resource scenarios while preserving language-specific phonological characteristics.
4.3.3. Data Augmentation Strategies
Addressing the persistent challenge of limited speech data, Boulal et al. [54] developed a comprehensive data augmentation framework specifically designed for Amazigh speech recognition. Their approach incorporated multiple augmentation techniques, including speed perturbation and pitch modification, while maintaining the phonological integrity of the source material. Experimental validation demonstrated statistically significant improvements in model generalization across varied acoustic conditions and dialectal variations.
Complementary work by Daouad et al. [55] extended these techniques through the introduction of context-aware augmentation methods. Their research indicated that the incorporation of linguistic knowledge into the augmentation process yields more naturalistic synthetic data, particularly beneficial for training robust recognition systems in low-resource scenarios.
Table 5 provides a comprehensive comparison of these various approaches and their respective performance metrics.
4.4. Challenges and Research Directions
Despite substantial methodological advancements, Amazigh speech technology continues to confront several significant challenges. Dialectal variation constitutes a persistent obstacle, with recognition performance exhibiting considerable variance across dialectal boundaries. Telmem et al. [10] conducted a systematic analysis of architectural performance across Amazigh dialects, revealing significant inter-dialectal performance differentials. These findings underscore the necessity of either dialect-specific modeling approaches or more sophisticated cross-dialectal architectures incorporating dialect-adaptive components.
The recognition of continuous speech presents additional complexities related to coarticulation effects and prosodic boundaries. Recent architectural innovations incorporating CTC loss functions [52] have demonstrated promise in addressing these challenges; however, performance metrics for connected speech recognition remain substantially below those achieved for isolated word recognition. This disparity highlights the need for continued research on temporal modeling approaches specifically calibrated for Amazigh’s phonological characteristics.
Resource constraints continue to limit research advancement despite innovative approaches to data augmentation and transfer learning. The development of larger, more diverse speech corpora remains a priority, particularly for under-represented dialects and demographic categories. The success of data augmentation approaches [54,55] suggests potential for synthetic data generation as a partial mitigation strategy; however, the acquisition of naturalistic speech data remains essential for robust system development.
Several promising research directions emerge from this analysis: Integration of attention mechanisms optimized for Amazigh phonological structures, particularly for capturing long-range dependencies in morphologically complex words. Development of unified modeling frameworks capable of cross-dialectal generalization, potentially through multitask learning approaches or dialect-adaptive layers. Application of semi-supervised and self-supervised learning paradigms to leverage unlabeled speech data, which typically exhibits greater availability than manually annotated resources. Domain adaptation methodologies for downstream applications, including educational technologies, accessibility solutions, and cultural heritage preservation.
These research trajectories collectively hold the potential to advance Amazigh speech technology while contributing to methodological innovations applicable to other under-resourced languages facing similar challenges.
5. Optical Character Recognition
Optical Character Recognition (OCR) systems for Amazigh script present unique computational challenges due to Tifinagh’s distinctive graphemic characteristics and the complexity of multi-script document processing. This section examines the methodological evolution from traditional recognition approaches to contemporary deep learning architectures across both handwritten and printed text domains.
5.1. Methodological Advances and Technical Innovations
Amazigh OCR systems have evolved through three distinct methodological phases (Table 6). The initial phase (2010–2015) was characterized by fundamental resource development and the application of classical pattern recognition techniques. Saady et al. [56] established the AMHCD database while Bencharef et al. [57] provided one of the first comprehensive datasets specifically designed for Tifinagh handwriting recognition. These resources established the essential infrastructure for empirical validation and comparative evaluation.
Supervised classification methods were systematically investigated by Aharrane et al. [58], whose comparative analysis revealed the differential efficacy of feature extraction techniques for Tifinagh recognition. Their implementation of multilayer perceptron architectures with optimized density and shadow features achieved 96.47% recognition accuracy on 24,180 handwritten characters [59], establishing performance benchmarks for subsequent research.
The intermediate phase (2016–2020) saw the development of end-to-end systems integrating multiple processing stages. Aharrane et al. [60] introduced a complete system for printed Amazigh script recognition, addressing the entire pipeline from preprocessing to character recognition. Particularly significant was their subsequent work on printed Tifinagh recognition in web images and natural scenes [61], which combined text region extraction with CNN-based language identification to achieve 99.12% accuracy in script identification and 99.93% in character recognition. This research demonstrated the feasibility of accurate Tifinagh recognition even in challenging multilingual environments characterized by complex backgrounds and variable illumination conditions.
The current phase (2021–2025) is characterized by architectural innovation and application to novel modalities. Chaabi [62] introduced a sophisticated hybrid CNN-LSTM architecture demonstrating significant performance improvements for both isolated and connected character recognition, while Rajaa and Ahmed [63] developed specialized convolutional neural networks addressing inter-writer variation and stylistic inconsistencies. Beyond traditional document processing, Rachidi [64] expanded OCR capabilities to video-captured text, introducing a system employing Random Forest classification for mobile video applications.
5.2. Technical Challenges and Solutions
Several persistent technical challenges have driven innovation in Amazigh OCR research. Character disambiguation represents a fundamental difficulty, particularly for Tifinagh graphemes exhibiting rotational or scalar similarities. While early approaches employed multiple descriptors and classifier ensembles, Erritali et al. [65] developed a more efficient search-based classification system, achieving a comparable recognition accuracy with reduced computational complexity.
Multi-script document processing presents additional complications, particularly for Amazigh-French documents prevalent in administrative and educational contexts. El Gajoui and Ataa Allah [66] systematically investigated these challenges, emphasizing the importance of robust script identification mechanisms and context-sensitive processing strategies. Complementary research by Gajoui et al. [67] addressed the processing of diacritical markers through neural network-based methodologies, demonstrating effective management of the complex relationship between base characters and associated diacritical elements.
The adaptation of existing OCR frameworks to Amazigh-specific requirements has produced promising results, as demonstrated by El Gajoui et al. [68] through their calibration of Tesseract for Tifinagh recognition. This research highlighted the potential for leveraging established technologies while addressing script-specific challenges, an approach particularly valuable in resource-constrained implementation scenarios.
Table 6Evolution of Amazigh OCR technologies (2010–2025).
| Recognition Task | 2010–2015 | 2016–2020 | 2021–2025 |
|---|---|---|---|
| Dataset Development | AMHCD database [56]; Dataset for handwriting recognition [57] | Dataset of printed words [69] | Video text datasets [64] |
| Handwritten Character Recognition | MLP with 96.47% accuracy [59]; Supervised classification methods [58] | Structural approach [70]; Statistical features [71] | CNN-LSTM architecture [62]; Specialized CNN [63]; ML algorithms [72] |
| Printed Text Recognition | Diacritical processing [67]; Tesseract adaptation [68]; Multilingual processing [66] | End-to-end document system [60]; Natural scene recognition [61] | Search-based classification [65]; Berber sign recognition [73] |
| Character Disambiguation | Multiple descriptors and classifiers | Statistical feature sets [71] | Search-based classification with 99.93% accuracy [65] |
| Alternative Modalities | Limited exploration | Initial video-based approaches | Random Forest for video text [64]; Mobile capture processing |
5.3. Implications and Research Directions
The advancement of Amazigh OCR technology has significant implications for digital preservation and accessibility of textual cultural heritage. Current systems have achieved impressive performance for both handwritten and printed text recognition, with accuracy rates exceeding 99% for certain tasks. However, several challenges persist that will shape future research directions.
The effective processing of degraded documents, the enhancement of recognition accuracy in adverse imaging conditions, and the development of more efficient training methodologies for limited data scenarios represent crucial areas for continued investigation. Sliman and Azouaoui [72] identify several promising directions for future research, including the integration of attention mechanisms for enhanced character localization, development of more robust preprocessing techniques, and enhancement of recognition accuracy through contextual information.
The expansion of OCR capabilities to non-traditional modalities, such as mobile video capture [64] and natural scene text [61], represents a significant advancement with practical implications for accessibility and heritage documentation. These developments demonstrate the potential of contemporary OCR technologies to support comprehensive digital preservation initiatives while accommodating diverse documentation scenarios.
The productive synthesis of traditional script analysis with contemporary computational methodologies continues to drive innovation in Amazigh OCR research. This interdisciplinary approach advances increasingly sophisticated solutions for the technical challenges inherent in Amazigh text recognition while contributing to the broader objectives of digital cultural heritage preservation and linguistic accessibility.
6. Machine Translation
Machine translation for the Amazigh language represents an emerging research domain characterized by significant methodological advancement despite persistent resource constraints. Recent developments have primarily focused on addressing the dual challenges of limited parallel data and substantial dialectal variation. This section examines the methodological evolution, resource development initiatives, and technical innovations that have shaped this nascent field.
6.1. Methodological Evolution and Resource Development
The development of Amazigh machine translation systems has progressed through distinct methodological phases, as illustrated in Table 7. Early research initiatives focused on establishing foundational frameworks and preliminary methodologies. This initial exploration phase transitioned to more structured approaches with Taghbalout et al.’s [9] investigation of Universal Networking Language (UNL) for Moroccan Amazigh translation. Their work established one of the first systematic methodologies, emphasizing the development of intermediate semantic representations capable of capturing Amazigh’s complex morphosyntactic features while facilitating interlingual transfer.
The field has witnessed significant methodological advancement with the integration of neural approaches. Maarouf et al. [36] demonstrated substantial progress in English-to-Amazigh translation through the application of transformer architectures. Their research directly addressed several fundamental challenges in Amazigh machine translation, particularly the handling of morphological complexity and cross-linguistic structural divergence. The implementation of transformer-based models represents a critical advancement beyond previous rule-based and statistical approaches, aligning Amazigh machine translation methodologies with contemporary techniques in the broader field.
Resource development has emerged as a crucial enabling factor for methodological advancement. The creation of the first parallel multi-lingual corpus of Amazigh by Ataa Allah and Miftah [35] established essential infrastructure for subsequent research. More recently, Maarouf’s [34] contribution of a comprehensive parallel corpus in Tifinagh-English represents a significant advancement in resource availability, providing researchers with valuable data for system training and evaluation.
6.2. Technical Challenges and Innovative Solutions
The persistent challenge of limited parallel data has stimulated innovative methodological approaches. Diab et al. [11] introduced guided back-translation techniques specifically optimized for the Kabyle-French language pair. Their research demonstrated how carefully structured guidance mechanisms can significantly enhance translation quality in low-resource scenarios. Complementary work by Lichouri and Abbas [74] addressed zero and low-resourced dialectal varieties through the utilization of an extended version of the dialectal parallel corpus (PADIC v2.0), demonstrating effective cross-dialectal transfer despite minimal training data.
Dialectal variation represents a persistent challenge for Amazigh machine translation systems. The linguistic divergence across Amazigh varieties necessitates specialized approaches for effective cross-dialectal processing. While initial research focused primarily on Central Atlas Tamazight, recent investigations have expanded to encompass additional varieties, notably Kabyle [11]. The development of effective cross-dialectal approaches remains an active research direction, with current methodologies typically employing dialect-specific models or transfer learning techniques to address inter-dialectal divergence.
The evaluation of Amazigh machine translation systems presents additional methodological challenges, particularly given the limited availability of standardized test sets and reference translations. Current research typically employs traditional metrics such as BLEU scores, though these may inadequately capture the linguistic nuances and morphological complexity of Amazigh. The development of specialized evaluation methodologies represents an important area for future investigation.
7. Resources and Datasets
The development and availability of specialized linguistic resources and structured datasets constitute a critical foundation for methodological advancement in Amazigh language processing. This section examines four principal categories of resources that have enabled significant technical progress across computational domains.
7.1. Linguistic Corpora and Annotations
The establishment of standardized linguistic corpora has proven instrumental to methodological advancement in Amazigh language processing research. Boulaknadel and Ataa Allah [33] formulated the foundational guidelines for constructing standardized Amazigh corpora, providing a methodological framework that continues to influence contemporary resource development initiatives. A significant advancement was achieved by Ataa Allah and Miftah [35], who introduced the first parallel multi-lingual corpus of Amazigh, establishing the essential infrastructure for cross-lingual research and machine translation system development.
Substantial contributions to morphosyntactic resources have emerged from multiple research initiatives. Amri et al. [20] developed specialized annotated corpora specifically designed for morphological and syntactic analysis, providing crucial data for training and evaluating computational models. This morphosyntactic foundation was subsequently enhanced by Bani et al. [37], who introduced the first Amazigh lemmatizer, demonstrating continued advancement in fundamental processing tools. Recent contributions by Faouzi et al. [38] in developing specialized corpora for word embedding evaluation have further augmented the available resources for distributional semantic analysis of Amazigh text.
7.2. Speech Recognition Databases
Speech technology research has been significantly facilitated through the development of specialized acoustic databases. El Ouahabi et al. [44] made a substantial contribution with their development of AMZSRD (Amazigh Speech Recognition Database), which has subsequently become a standard resource for speech recognition system development and evaluation. This comprehensive database incorporates multiple variables, including speaker demographic diversity, dialectal variations, and controlled recording conditions, enabling systematic investigation of these parameters’ effects on recognition performance.
The significance of these acoustic resources is evidenced in contemporary research initiatives. Boulal et al. [54] leveraged such databases for the development and evaluation of sophisticated data augmentation strategies designed to mitigate data scarcity constraints. Similarly, Telmem et al. [10] demonstrated the methodological value of well-structured speech datasets in their systematic comparative analysis of different recognition architectures across demographic variables. These applications highlight the critical role of specialized speech resources in enabling rigorous experimental evaluation and methodological advancement.
7.3. Character Recognition Datasets
Optical Character Recognition research has been facilitated by dedicated datasets for both handwritten and printed Tifinagh graphemes. Initial contributions in this domain came from Gajoui et al. [76], who developed a specialized corpus for Latin OCR systems evaluation, particularly significant for transcription system development. This foundational work was complemented by Aharrane et al. [69], who established a comprehensive dataset of Amazigh printed words images that has become essential infrastructure for OCR system development and evaluation.
Subsequent advances in character recognition resources include Bencharef et al.’s [57] specialized dataset for Tifinagh handwriting recognition and Saady et al.’s [56] development of the AMHCD database, which has become a standard resource for handwritten character recognition research. These complementary datasets have enabled the systematic evaluation of recognition methodologies, as demonstrated by Aharrane et al. [58] in their rigorous comparative analysis of supervised classification methods. The diversity of these resources across both handwritten and printed domains has been instrumental in addressing the script-specific challenges of Tifinagh recognition.
7.4. Translation Resources
Recent research initiatives have increasingly focused on developing specialized resources for machine translation applications. Maarouf [34] made a significant contribution with the development of a comprehensive linguistic dataset incorporating a parallel corpus in Tifinagh-English. This resource addresses a critical gap in available translation data, particularly for contemporary neural machine translation approaches that require substantial parallel text for effective training.
The development of dialectally diverse parallel corpora has been advanced by Lichouri and Abbas [74] through their work on the dialectal parallel corpus (PADIC v2.0), providing valuable resources for translation between Arabic dialects and Amazigh variants. These specialized translation resources have proven essential for developing and evaluating translation systems, as evidenced in recent methodological investigations by Diab et al. [11] on guided back-translation techniques for low-resource language pairs.
Collectively, these diverse linguistic resources constitute the essential infrastructure supporting methodological advancement across Amazigh language processing domains. Their continued development, standardization, and accessibility represent critical factors in the field’s ongoing progress. Future resource development initiatives will likely focus on addressing the remaining gaps in dialectal coverage, domain diversity, and semantic annotation, while increasing the scale and quality of existing resource types.
7.5. Resource Impact on Technological Development
The relationship between resource development and technological advancement in Amazigh language processing demonstrates clear dependencies. Table 8 illustrates how specific resources have enabled key technological breakthroughs across domains.
Three development phases demonstrate this resource-technology relationship:
Foundation (2010–2015): Character datasets enabled first OCR systems; basic lexicons supported rule-based morphological analyzers achieving >95% accuracy.
Enhancement (2016–2020): Annotated corpora enabled statistical ML approaches; speech databases facilitated transition from HMM to neural architectures; parallel corpora enabled systematic MT research.
Neural Integration (2021–2025): Large-scale parallel corpora enabled transformer approaches; augmented datasets improved model generalization across acoustic conditions and dialectal variations.
This progression illustrates that strategic resource development, coordinated with computational requirements, has been the primary driver of technological advancement across all Amazigh processing domains.
8. Applications and Impact
The development of Amazigh language technologies has yielded substantive applications with significant societal implications, particularly in educational contexts, cultural preservation initiatives, and digital accessibility domains. This section examines the practical implementations and broader impact of these technological advancements.
8.1. Educational and Cultural Preservation Applications
The implementation of language technologies in educational contexts represents a particularly significant application domain. Guerchouh and Suçin [75] conducted a systematic examination of artificial intelligence applications in language learning, specifically addressing the role of machine translation in Tamazight educational contexts. Their analysis demonstrated how automated translation systems can effectively augment both pedagogical delivery and learner engagement. This educational integration has been complemented by El Ouahabi et al.’s [77] evaluation of speech recognition technologies for pronunciation training and oral practice, revealing significant enhancement in learning outcomes through interactive digital tools.
The digital preservation of cultural heritage has emerged as a domain of particular significance. Corallo and Varde [73] demonstrated the application of optical character recognition technologies for the digitization and preservation of historical Amazigh texts, ensuring both the accessibility and long-term conservation of these cultural materials. The methodological frameworks established by Boulaknadel and Ataa Allah [33] for standardized corpus development have enabled systematic documentation of dialectal variations and linguistic forms, creating digital archives of substantial cultural significance.
The application of language technologies to literary and religious texts represents an additional dimension of cultural impact. Klaassen et al. [78] documented the methodological challenges and cultural implications of Biblical translation for indigenous Islamic communities, illustrating how translation technologies can support culturally sensitive rendering of religious texts. Complementary work by El Gajoui and Ataa Allah [66] on multilingual document processing has facilitated the digitization and dissemination of Amazigh literary works, preserving their linguistic characteristics while expanding their accessibility.
8.2. Digital Accessibility and Research Impact
The advancement of language processing tools has substantially enhanced digital accessibility for Amazigh speakers. Maarouf et al.’s [36] development of automatic translation systems has addressed the critical communication barriers between Amazigh speakers and other language communities. These technologies have demonstrated particular significance for online content accessibility, government service provision, educational resource availability, and cross-cultural communication contexts.
The research impact of these technological developments extends beyond immediate applications to encompass broader linguistic investigation. Ataa Allah and Boulaknadel [39] conducted a comprehensive analysis of technological trends in less-resourced language processing, demonstrating how computational tools have enabled more systematic documentation and analysis of Amazigh dialectal variations. The development of standardized resources by Amri et al. [20] has similarly facilitated more rigorous linguistic research methodologies, providing essential infrastructure for investigating dialectal variation and linguistic evolution.
The broader societal implications of these technologies include significant contributions to language revitalization and cultural preservation initiatives. The creation of comprehensive linguistic datasets and computational tools, exemplified by Maarouf’s [34] recent contributions, provides the foundational infrastructure for expanding Amazigh language use in digital contexts. These developments collectively represent a significant advancement in digital language equity, ensuring that historically marginalized languages can participate fully in contemporary technological ecosystems.
As these technologies continue to develop, future applications will likely focus on enhancing cross-dialectal access, expanding domain coverage beyond current limitations, and deepening the integration with educational systems. The convergence of multiple technological domains—from speech recognition to machine translation—suggests opportunities for increasingly sophisticated applications that address the specific needs of Amazigh language communities within both educational and broader societal contexts.
9. Discussion, Challenges, and Future Research Directions
9.1. Summary of Achievements
The development of Amazigh language technologies over the past fifteen years has demonstrated both significant achievements and persistent challenges. The field has evolved from isolated research efforts to a cohesive domain with substantial methodological advancements across multiple areas. The transition from rule-based to neural approaches represents a thoughtful adaptation that addresses Amazigh’s unique linguistic characteristics.
A critical factor in this evolution has been the synergistic relationship between resource development and methodological innovation. Foundational resources, such as morphosyntactically annotated corpora and specialized speech databases, have enabled the sophisticated neural approaches now prevalent in the field. Current capabilities suggest that Amazigh language technology has reached a significant inflection point, with functional processing tools across all major domains enabling practical applications with viable performance levels.
9.2. Persistent Challenges
Resource Limitations: Critical gaps persist with machine translation requiring 1 M+ parallel sentences versus 50 K available, speech recognition needing 1000+ h versus 100 h accessible, and NER systems requiring 50 K+ annotated entities compared to 10 K existing annotations.
Dialectal Complexity: Resource allocation heavily favors Central Atlas Tamazight (60%), creating 15–30% accuracy degradation for under-resourced dialects. This imbalance may exacerbate digital divides within Amazigh communities.
Morphological Processing: Amazigh’s templatic morphology with extensive affixation exceeds the complexity of many well-resourced languages. Neural models struggle with unseen morphological patterns and suffer from data sparsity.
Evaluation Limitations: Standard metrics inadequately capture Amazigh’s complexity. BLEU scores fail to account for morphological variants and dialectal differences, while WER does not distinguish between dialectal pronunciation variants and actual errors.
9.3. Future Research Directions
Technical Innovations: Cross-dialectal models using multi-task learning frameworks, advanced data augmentation through synthetic generation, integration of attention mechanisms optimized for Amazigh phonological structures, and development of Amazigh-specific evaluation metrics.
Emerging Opportunities: Large language model integration through few-shot learning and cross-lingual transfer, multimodal approaches combining vision-language and speech-text fusion, and community-driven development through participatory design and gamified data collection.
Application Priorities: Educational technologies including adaptive language learning systems, digital accessibility through government service translation, and cultural heritage preservation through comprehensive digitization initiatives.
The successful development of language technologies for Amazigh represents more than a technical achievement—it constitutes a significant advancement toward digital language equity. The field’s continued progress depends on a sustained commitment to creating inclusive technologies that serve all speakers while advancing broader low-resource language processing methodologies.
Conceptualization, O.A. and K.F.; methodology, O.A.; investigation, O.A. and M.A.; resources, O.A. and M.A.; data curation, O.A.; writing—original draft preparation, O.A.; writing—review and editing, M.A. and K.F.; visualization, O.A.; supervision, K.F.; project administration, O.A. All authors have read and agreed to the published version of the manuscript.
The authors declare no conflicts of interest.
Footnotes
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Figure 1 Geographic distribution of Amazigh language varieties across North Africa, showing major dialect regions and speaker populations.
Figure 2 The Tifinagh alphabet standardized by IRCAM for Modern Amazigh writing.
Figure 3 Evolution of Amazigh language technologies (2010–2025), demonstrating the transition from rule-based approaches through statistical methods to neural architectures.
State of Amazigh language technology development (2025).
| Technology | Maturity | Resources | Performance | Dialect | Research |
|---|---|---|---|---|---|
| Domain | Level | Available | Metrics | Coverage | Activity |
| Morphological Analysis | High | Med | 95%+ acc. | Wide | High |
| POS Tagging | High | Med | 93.8% acc. | Med | High |
| Named Entity Recognition | Med | Low | 82% F1 | Limited | Med |
| Speech Recognition | High | Med | 96% acc. | Med | High |
| (Isolated Words) | |||||
| Speech Recognition | Med | Low | 73% acc. | Limited | Med |
| (Continuous) | |||||
| OCR (Printed) | High | High | 99.9% acc. | Wide | Med |
| OCR (Handwritten) | Med | Med | 93% acc. | Med | High |
| Machine Translation | Low | Low | BLEU: 14–22 | Limited | High |
Maturity Level: Technical readiness (High/Med/Low); Resources: Available datasets and tools; Performance: Best reported results from literature; Dialect Coverage: Support across Amazigh varieties; Research Activity: Current research publications.
Resource distribution across Amazigh dialects.
| Dialect | Speech | Text | MT | Performance Impact |
|---|---|---|---|---|
| Central Atlas Tamazight | High | High | Medium | Baseline performance |
| Tarifit | Medium | Medium | Low | 15–20% accuracy drop |
| Tachelhit | Medium | Low | Low | 20–25% accuracy drop |
| Kabyle | Medium | Medium | Medium | 10–15% accuracy drop |
| Other varieties | Low | Very Low | Very Low | 25–30% accuracy drop |
Distribution of publications across Amazigh language technology domains.
| Research Domain | Number of Papers | Percentage | Time Period Focus |
|---|---|---|---|
| Natural Language Processing | 25 | 35% | 2012–2025 |
| Speech Technologies | 22 | 31% | 2014–2025 |
| Optical Character Recognition | 18 | 25% | 2011–2024 |
| Machine Translation | 5 | 7% | 2018–2025 |
| Resources & Datasets | 2 | 3% | 2012–2025 |
| Total | 72 | 100% | 2010–2025 |
Evolution of Amazigh NLP technologies (2010–2025).
| Task | 2010–2015 | 2016–2020 | 2021–2025 |
|---|---|---|---|
| Morphological Analysis | Rule-based systems [ | Finite-state models [ | Neural-enhanced approaches; dialect adaptations [ |
| POS Tagging | Tagset design [ | Statistical ML [ | LSTM, GRU, DNN architectures (93.8% accuracy) [ |
| Named Entity Recognition | Rule-based frameworks [ | SVM, hybrid systems [ | Enhanced hybrid methods [ |
| Semantic Processing | Basic syntax analysis [ | Syntactic-semantic integration [ | Limited advancement |
| Resources | Initial corpora [ | Annotated data [ | Word embeddings, lemmatizer, comprehensive datasets [ |
| Translation | Limited exploration | Parallel corpora [ | Transformer MT, back-translation [ |
Comparison of Amazigh speech recognition approaches.
| Study | Architecture | Features | Dataset | Performance |
|---|---|---|---|---|
| [ | HMM-based | MFCC | Small vocabulary | Preliminary results |
| [ | HMM, GMM | MFCC | 187 words, 50 speakers | 8.20% WER |
| [ | CNN | MFCC | 9240 files, gender analysis | 93.9% accuracy |
| [ | CNN | MFCC | Digit recognition, 42 speakers | 91.75% accuracy |
| [ | 1D & 2D CNN-LSTM | Spectrogram | 2400 files, 80 speakers | >96% accuracy |
| [ | CNN-LSTM | Spectrogram | Gender & age corpus | Improved over CNN |
| [ | CNN-transformer | Mel-spectrogram | Tarifit corpus | Connected speech improvement |
| [ | Fine-tuned Whisper | Multiple | Limited samples | Effective transfer learning |
| [ | CNN-CTC | Multiple | Mid-sized corpus | Efficiency-accuracy balance |
| [ | CNN + augmentation | Mel-spectrogram | Augmented corpus | Better generalization |
Evolution of Amazigh machine translation approaches (2010–2025).
| Aspect | 2010–2015 | 2016–2020 | 2021–2025 |
|---|---|---|---|
| Methodological Approaches | Initial exploration; Rule-based systems | UNL-based approaches [ | Transformer architectures [ |
| Resource Development | Limited bilingual lexicons; Small-scale dictionaries | First parallel multi-lingual corpus [ | Tifinagh-English parallel corpus [ |
| Dialect-Specific Adaptation | Primarily focused on Central Atlas Tamazight | Initial multi-dialect exploration | Kabyle-French pair [ |
| Application Domains | Theoretical frameworks; Basic systems | Lexicon-driven translation; Cultural content | Educational applications [ |
Key resources and their technological impact.
| Resource | Domain | Impact | Key Achievement Enabled |
|---|---|---|---|
| AMHCD Database [ | OCR | Critical | First handwritten recognition (96.47% acc.) |
| AMZSRD Speech DB [ | ASR | Critical | CNN-LSTM architectures (>96% acc.) |
| Morphosyntactic Corpus [ | POS Tagging | High | Neural POS tagging (93.8% acc.) |
| First Parallel Corpus [ | MT | Foundational | Systematic MT research |
| Tifinagh-English Corpus [ | MT | Emerging | Transformer-based translation |
1. Joshi, P.; Santy, S.; Budhiraja, A.; Bali, K.; Choudhury, M. The state and fate of linguistic diversity and inclusion in the NLP world. arXiv; 2020; arXiv: 2004.09095
2. Aissati, A.E.; Karsmakers, S.; Kurvers, J. ‘We are all beginners’: Amazigh in language policy and educational practice in Morocco. Comp. J. Comp. Int. Educ.; 2011; 41, pp. 211-227. [DOI: https://dx.doi.org/10.1080/03057925.2011.547289]
3. Ait Laaguid, B.; Khaloufi, A. Amazigh language use on social media: An exploratory study. J. Arbitrer; 2023; 10, pp. 24-34. [DOI: https://dx.doi.org/10.25077/ar.10.1.24-34.2023]
4. Outahajala, M.; Zenkouar, L.; Rosso, P.; Martí, A. Tagging amazigh with ancorapipe. Proceedings of the Workshop on Language Resources and Human Language Technology for Semitic Languages; Valletta, Malta, 26 January 2010; pp. 52-56.
5. Galla, C.K. Indigenous language revitalization, promotion, and education: Function of digital technology. Comput. Assist. Lang. Learn.; 2016; 29, pp. 1137-1151. [DOI: https://dx.doi.org/10.1080/09588221.2016.1166137]
6. Fadoua, A.A.; Siham, B. Natural language processing for Amazigh language: Challenges and future directions. Lang. Technol. Norm. Less-Resour. Lang.; 2012; 19, 23.
7. Outahajala, M. Processing Amazighe Language. Proceedings of the Natural Language Processing and Information Systems: 16th International Conference on Applications of Natural Language to Information Systems, NLDB 2011; Alicante, Spain, 28–30 June 2011; Proceedings 16 Springer: Berlin/Heidelberg, Germany, 2011; pp. 313-317.
8. Allah, F.A.; Boulaknadel, S. Toward computational processing of less resourced languages: Primarily experiments for Moroccan Amazigh language. Theory and Applications for Advanced Text Mining; IntechOpen: London, UK, 2012.
9. Taghbalout, I.; Allah, F.A.; Marraki, M.E. Towards UNL-based machine translation for Moroccan Amazigh language. Int. J. Comput. Sci. Eng.; 2018; 17, pp. 43-54. [DOI: https://dx.doi.org/10.1504/IJCSE.2018.094418]
10. Telmem, M.; Laaidi, N.; Ghanou, Y.; Hamiane, S.; Satori, H. Comparative study of CNN, LSTM and hybrid CNN-LSTM model in amazigh speech recognition using spectrogram feature extraction and different gender and age dataset. Int. J. Speech Technol.; 2024; 27, pp. 1121-1133. [DOI: https://dx.doi.org/10.1007/s10772-024-10154-0]
11. Diab, N.; Sadat, F.; Semmar, N. Towards Guided Back-translation for Low-resource languages-A Case Study on Kabyle-French. Proceedings of the 2024 16th International Conference on Human System Interaction (HSI); Paris, France, 8–11 July 2024; pp. 1-4.
12. Raiss, H.; Cavalli-Sforza, V. ANMorph: Amazigh nouns morphological analyzer. Proceedings of the 5th International Conference on Amazigh and ICT; Rabat, Morocco, 26–27 November 2012.
13. Nejme, F.Z.; Boulaknadel, S.; Aboutajdine, D. Finite state morphology for Amazigh language. Proceedings of the International Conference on Intelligent Text Processing and Computational Linguistics; Samos, Greece, 24–30 March 2013; Springer: Berlin/Heidelberg, Germany, 2013; pp. 189-200.
14. Nejme, F.Z.; Boulaknadel, S.; Aboutajdine, D. AmAMorph: Finite state morphological analyzer for amazighe. J. Comput. Inf. Technol.; 2016; 24, pp. 91-110. [DOI: https://dx.doi.org/10.20532/cit.2016.1002478]
15. Ammari, R.; Zenkoua, A. APMorph: Finite-state transducer for Amazigh pronominal morphology. Int. J. Electr. Comput. Eng.; 2021; 11, 699. [DOI: https://dx.doi.org/10.11591/ijece.v11i1.pp699-706]
16. Oussou, S. Amazigh Language Morphology: Examples from Tashlhiyt in Ayt Hdidou. Lingua. Lang. Cult.; 2021; 20, pp. 212-224.
17. Outahajala, M.; Zenkouar, L.; Benajiba, Y.; Rosso, P. The development of a fine grained class set for Amazigh POS tagging. Proceedings of the 2013 ACS International Conference on Computer Systems and Applications (AICCSA); Ifrane, Morocco, 27–30 May 2013; IEEE: New York, NY, USA, 2013; pp. 1-8.
18. Samir, A.; Lahbib, Z.; Mohamed, O. Amazigh PoS tagging using machine learning techniques. Proceedings of the 2nd Mediterranean Symposium on Smart City Applications; Tangier, Morocco, 15–27 October 2017; pp. 551-562.
19. Amri, S.; Zenkouar, L. Amazigh POS tagging using TreeTagger: A language independant model. Proceedings of the Advanced Intelligent Systems for Sustainable Development (AI2SD’2018); Tangiers, Morocco, 12–14 July 2018; pp. 622-632.
20. Amri, S.; Zenkouar, L.; Outahajala, M. Build a morphosyntaxically annotated amazigh corpus. Proceedings of the 2nd International Conference on Big Data, Cloud and Applications; Tetouan, Morocco, 29–30 March 2017; pp. 1-7.
21. Maarouf, O.; El Ayachi, R. Part-of-Speech Tagging Using Long Short Term Memory (LSTM): Amazigh Text Written in Tifinaghe Characters. Proceedings of the 6th International Conference on Business Intelligence; Beni Mellal, Morocco, 27–29 May 2021; Springer: Berlin/Heidelberg, Germany, 2021; pp. 3-17.
22. Otman, M.; Mohamed, B. Amazigh part of speech tagging using gated recurrent units (GRU). Proceedings of the 2021 7th International Conference on Optimization and Applications (ICOA); Wolfenbüttel, Germany, 19–20 May 2021; IEEE: New York, NY, USA, 2021; pp. 1-6.
23. Bani, R.; Amri, S.; Zenkouar, L.; Guennoun, Z. Deep neural networks for part-of-speech tagging in under-resourced Amazigh. Rev. D’Intelligence Artif.; 2023; 37, 611. [DOI: https://dx.doi.org/10.18280/ria.370310]
24. Amri, S.; Zenkouar, L.; Benkhouya, R. A comparative study on the efficiency of POS tagging techniques on Amazigh corpus. Proceedings of the NISS19: Networking, Information Systems & Security; Rabat, Morocco, 27–29 March 2019; pp. 1-5.
25. Boulaknadel, S.; Talha, M.; Aboutajdine, D. Amazighe Named Entity Recognition using a A rule based approach. Proceedings of the 2014 IEEE/ACS 11th International Conference on Computer Systems and Applications (AICCSA); Doha, Qatar, 10–13 November 2014; IEEE: New York, NY, USA, 2014; pp. 478-484.
26. Talha, M.; Boulaknadel, S.; Aboutajdine, D. NERAM: Named Entity Recognition for AMazigh language (RENAM: Système de Reconnaissance des Entités Nommées Amazighes) [in French]. Proceedings of the TALN 2014 (Volume 2: Short Papers); Association Pour le Traitement Automatique des Langues: Marseille, France, 2014; pp. 517-524.
27. Talha, M.; Boulaknadel, S.; Aboutajdine, D. Performance Evaluation of SVM-Based Amazighe Named Entity Recognition. Proceedings of the International Conference on Advanced Machine Learning Technologies and Applications (AMLTA2018); Cairo, Egypt, 22–24 February 2018; Springer: Berlin/Heidelberg, Germany, 2018; pp. 232-241.
28. Talha, M.; Boulaknadel, S.; Aboutajdine, D. Development of Amazighe Named Entity Recognition System Using Hybrid Method. Res. Comput. Sci.; 2015; 90, pp. 151-161. [DOI: https://dx.doi.org/10.13053/rcs-90-1-11]
29. Talha, M.; Boulaknadel, S.; Aboutajdine, D. Enhancing Performance of Hybrid Named Entity Recognition for Amazighe Language. Machine Learning Paradigms: Theory and Application; Springer: Cham, Switzerland, 2019; pp. 211-232.
30. Amri, S.; Zenkouar, L.; Benkhouya, R. A hybrid statistical approach for named entity recognition for amazighe language. Proceedings of the 4th International Conference on Big Data and Internet of Things; Rabat, Morocco, 23–24 October 2019; pp. 1-6.
31. Nejme, F.Z.; Boulaknadel, S.; Aboutajdine, D. Toward an amazigh language processing. Proceedings of the 3rd Workshop on South and Southeast Asian Natural Language Processing; Mumbai, India, 8 December 2012; pp. 173-180.
32. Meryem, T. Analyse Syntactico-Sémantique Automatique de la Langue Amazighe. 2019; Available online: https://toubkal.imist.ma/handle/123456789/15334 (accessed on 12 March 2025).
33. Boulaknadel, S.; Ataa Allah, F. Building a standard Amazigh corpus. Proceedings of the Third International Conference on Intelligent Human Computer Interaction (IHCI 2011); Prague, Czech Republic, 29–31 August 2011; Springer: Cham, Switzerland, 2012; pp. 91-98.
34. Maarouf, O. Amazigh Linguistic Dataset: Part-of-Speech Tagging, Named Entity Recognition, and Parallel Corpus (Tifinagh-English). 2025; Version 1 Available online: https://data.mendeley.com/datasets/vdgfhfnr26/1 (accessed on 3 February 2025).
35. Allah, F.A.; Miftah, N. The First Parallel Multi-lingual Corpus of Amazigh. Fadoua ATAA ALLAH J. Eng. Res. Appl.; 2018; 8, pp. 5-12.
36. Maarouf, O.; Maarouf, A.; El Ayachi, R.; Biniz, M. Automatic translation from English to Amazigh using transformer learning. Indones. J. Electr. Eng. Comput. Sci.; 2024; 34, pp. 1924-1934. [DOI: https://dx.doi.org/10.11591/ijeecs.v34.i3.pp1924-1934]
37. Bani, R.; Amri, S.; Zenkouar, L.; Guennoun, Z. Amazlem: The First Amazigh Lemmatizer. Proceedings of the International Conference on Advanced Research in Technologies, Information, Innovation and Sustainability; Madrid, Spain, 18–20 October 2023; Springer: Cham, Switzerland, 2023; pp. 375-385.
38. Faouzi, H.; El-Badaoui, M.; Boutalline, M.; Tannouche, A.; Ouanan, H. Towards amazigh word embedding: Corpus creation and word2vec models evaluations. Rev. D’Intelligence Artif.; 2023; 37, 753. [DOI: https://dx.doi.org/10.18280/ria.370324]
39. Ataa Allah, F.; Boulaknadel, S. New trends in less-resourced language processing: Case of Amazigh language. Int. J. Nat. Lang. Comput. (IJNLC); 2023; 12, Available online: https://ssrn.com/abstract=4443059 (accessed on 12 March 2025).
40. Daouad, M.; Allah, F.A.; Dadi, E.W. An automatic speech recognition system for isolated Amazigh word using 1D & 2D CNN-LSTM architecture. Int. J. Speech Technol.; 2023; 26, pp. 775-787.
41. Satori, H.; ElHaoussi, F. Investigation Amazigh speech recognition using CMU tools. Int. J. Speech Technol.; 2014; 17, pp. 235-243. [DOI: https://dx.doi.org/10.1007/s10772-014-9223-y]
42. El Ouahabi, S.; Atounti, M.; Bellouki, M. Toward an automatic speech recognition system for amazigh-tarifit language. Int. J. Speech Technol.; 2019; 22, pp. 421-432. [DOI: https://dx.doi.org/10.1007/s10772-019-09617-6]
43. Telmem, M.; Ghanou, Y. Estimation of the optimal HMM parameters for amazigh speech recognition system using CMU-Sphinx. Procedia Comput. Sci.; 2018; 127, pp. 92-101. [DOI: https://dx.doi.org/10.1016/j.procs.2018.01.102]
44. El Ouahabi, S.; Atounti, M.; Bellouki, M. A database for Amazigh speech recognition research: AMZSRD. Proceedings of the 2017 3rd International Conference of Cloud Computing Technologies and Applications (CloudTech); Rabat, Morocco, 24–26 October 2017; IEEE: New York, NY, USA, 2017; pp. 1-5.
45. Telmem, M.; Ghanou, Y. A comparative study of HMMs and CNN acoustic model in amazigh recognition system. Proceedings of the Embedded Systems and Artificial Intelligence: Proceedings of ESAI 2019; Fez, Morocco, 2–3 May 2019; Springer: Cham, Switzerland, 2020; pp. 533-540.
46. Telmem, M.; Ghanou, Y. The convolutional neural networks for Amazigh speech recognition system. Telkomnika (Telecommun. Comput. Electron. Control); 2021; 19, pp. 515-522. [DOI: https://dx.doi.org/10.12928/telkomnika.v19i2.16793]
47. Hamidi, M.; Satori, H.; Zealouk, O.; Satori, K. Amazigh digits through interactive speech recognition system in noisy environment. Int. J. Speech Technol.; 2020; 23, pp. 101-109. [DOI: https://dx.doi.org/10.1007/s10772-019-09661-2]
48. Boulal, H.; Hamidi, M.; Abarkan, M.; Barkani, J. Amazigh spoken digit recognition using a deep learning approach based on mfcc. Int. J. Electr. Comput. Eng. Syst.; 2023; 14, pp. 791-798. [DOI: https://dx.doi.org/10.32985/ijeces.14.7.6]
49. Telmem, M.; Laaidi, N.; Satori, H. The impact of MFCC, spectrogram, and Mel-Spectrogram on deep learning models for Amazigh speech recognition system. Int. J. Speech Technol.; 2025; 28, pp. 299-312. [DOI: https://dx.doi.org/10.1007/s10772-025-10183-3]
50. Boulal, H.; Hamidi, M.; Abarkan, M.; Barkani, J. Amazigh CNN speech recognition system based on Mel spectrogram feature extraction method. Int. J. Speech Technol.; 2024; 27, pp. 287-296. [DOI: https://dx.doi.org/10.1007/s10772-024-10100-0]
51. Daouad, M.; Allah, F.A.; Dadi, E.W. Amazigh speech recognition via parallel CNN transformer-encoder model. Proceedings of the International Conference on Digital Age & Technological Advances for Sustainable Development; Kosice, Slovakia, 27–29 May 2024; Springer: Cham, Switzerland, 2024; pp. 255-263.
52. Telmem, M. Build an efficient Amazigh Speech Recognition using a CNN-CTC model. Proceedings of the 2024 3rd International Conference on Embedded Systems and Artificial Intelligence (ESAI); Fez, Morocco, 19–20 December 2024; IEEE: New York, NY, USA, 2024; pp. 1-7.
53. Daouad, M.; Ataa Allah, F.; Dadi, E.W. Optimizing Whisper models for Amazigh ASR: A comparative analysis. Int. J. Speech Technol.; 2024; 28, pp. 27-37. [DOI: https://dx.doi.org/10.1007/s10772-024-10158-w]
54. Boulal, H.; Bouroumane, F.; Hamidi, M.; Barkani, J.; Abarkan, M. Exploring data augmentation for Amazigh speech recognition with convolutional neural networks. Int. J. Speech Technol.; 2025; 28, pp. 53-65. [DOI: https://dx.doi.org/10.1007/s10772-024-10164-y]
55. Daouad, M.; Ataa Allah, F.; Dadi, E.W. Enhancing Automatic Speech Recognition Systems for Amazigh Language Through Data Augmentation. Proceedings of the International Conference On Big Data and Internet of Things; Macau, China, 14–16 September 2024; Springer: Cham, Switzerland, 2024; pp. 866-879.
56. Saady, Y.E.; Rachidi, A.; Yassa, M.; Mammass, D. Amhcd: A database for amazigh handwritten character recognition research. Int. J. Comput. Appl.; 2011; 27, pp. 44-48.
57. Bencharef, O.; Chihab, Y.; Mousaid, N.; Oujaoura, M. Data set for Tifinagh handwriting character recognition. Data Brief; 2015; 4, 11. [DOI: https://dx.doi.org/10.1016/j.dib.2015.04.008]
58. Aharrane, N.; El Moutaouakil, K.; Satori, K. A comparison of supervised classification methods for a statistical set of features: Application: Amazigh OCR. Proceedings of the 2015 Intelligent Systems and Computer Vision (ISCV); Fez, Morocco, 25–26 March 2015; IEEE: New York, NY, USA, 2015; pp. 1-8.
59. Aharrane, N.; El Moutaouakil, K.; Satori, K. Recognition of handwritten Amazigh characters based on zoning methods and MLP. WSEAS Trans. Comput.; 2015; 14, pp. 178-185.
60. Aharrane, N.; Dahmouni, A.; Ensah, K.E.M.; Satori, K. End-to-end system for printed Amazigh script recognition in document images. Proceedings of the 2017 International Conference on Advanced Technologies for Signal and Image Processing (ATSIP); Fez, Morocco, 22–24 May 2017; IEEE: New York, NY, USA, 2017; pp. 1-6.
61. Aharrane, N.; Dahmouni, A.; El Moutaouakil, K.; Satori, K. Printed Tifinagh script recognition from Web and natural scenes images in multilingual environment. Proceedings of the TICAM 2018: International Conference on Information and Communication Technologies for Amazigh; Rabat, Morocco, 26–27 November 2018.
62. Chaabi, Y. Tifinagh Characters Recognition Using Deep CNN-LSTM. Proceedings of the 2024 4th International Conference on Emerging Smart Technologies and Applications (eSmarTA); Sana’a, Yemen, 6–7 August 2024; IEEE: New York, NY, USA, 2024; pp. 1-8.
63. Rajaa, S.; Ahmed, A. Tifinagh Handwritten Character Recognition Using the Convolutional Neural Network. Proceedings of the 2024 11th International Conference on Wireless Networks and Mobile Communications (WINCOM); Leeds, UK, 23–25 July 2024; IEEE: New York, NY, USA, 2024; pp. 1-5.
64. Rachidi, Y. Random Forest for video Text Amazigh. Proceedings of the E3S Web of Conferences; Kryvyi Rih, Ukraine, 19–21 May 2021; EDP Sciences: Les Ulis, France, 2021; Volume 229, 01062.
65. Erritali, M.; Chouni, Y.; Ouadid, Y. Search-Based Classification for Offline Tifinagh Alphabets Recognition. Advancements in Computer Vision Applications in Intelligent Systems and Multimedia Technologies; IGI Global: Hershey, PA, USA, 2020; pp. 255-267.
66. El Gajoui, K.; Allah, F.A. Optical character recognition for multilingual documents: Amazigh-French. Proceedings of the 2014 Second World Conference on Complex Systems (WCCS); Agadir, Morocco, 10–12 November 2014; IEEE: New York, NY, USA, 2014; pp. 84-89.
67. Gajoui, K.E.; Allah, F.A.; Oumsis, M. Diacritical Language OCR based on neural network: Case of Amazigh language. Procedia Comput. Sci.; 2015; 73, pp. 298-305. [DOI: https://dx.doi.org/10.1016/j.procs.2015.12.035]
68. el Gajoui, K.; Allah, F.A.; Oumsis, M. Training tesseract tool for amazigh ocr. Proceedings of the Recent Researches in Applied Computer Science: Proceedings of the 15 th International Conference on Applied Computer Science; Konya, Turkey, 20–22 May 2015; pp. 20-22.
69. Aharrane, N.; Dahmouni, A.; EL MOUTAOUAKIL, K.; Satori, K. A Dataset of Amazigh Printed Words Images. Proceedings of the TICAM 2016: International Conference on Information and Communication Technologies for Amazigh; Rabat, Morocco, 26–27 November 2016.
70. Ouadid, Y.; Minaoui, B.; Elbalaoui, A.; Fakir, M.; Ahdid, R. Handwritten Tifinagh Character Recognition Through a Structural Approach. Proceedings of the 2017 14th International Conference on Computer Graphics, Imaging and Visualization; Marrakesh, Morocco, 23–25 May 2017; IEEE: New York, NY, USA, 2017; pp. 49-55.
71. Aharrane, N.; Dahmouni, A.; El Moutaouakil, K.; Satori, K. A robust statistical set of features for Amazigh handwritten characters. Pattern Recognit. Image Anal.; 2017; 27, pp. 41-52. [DOI: https://dx.doi.org/10.1134/S1054661817010011]
72. Sliman, R.; Azouaoui, A. Tifinagh Handwritten Character Recognition Using Machine Learning Algorithms. International Conference on Advanced Intelligent Systems for Sustainable Development; Springer: Berlin/Heidelberg, Germany, 2022; pp. 26-34.
73. Corallo, L.; Varde, A.S. Optical Character Recognition and Transcription of Berber Signs from Images in a Low-Resource Language Amazigh. arXiv; 2023; arXiv: 2303.13549
74. Lichouri, M.; Abbas, M. Machine translation for zero and low-resourced dialects using a new extended version of the dialectal parallel corpus (Padic v2. 0). Proceedings of the 4th International Conference on Natural Language and Speech Processing (ICNLSP 2021); Trento, Italy, 12–13 November 2021; pp. 33-38.
75. Lydia, G.; Suçin, M.H. Impact de l’intelligence artificielle (la traduction automatique) dans l’apprentissage des langues: Cas du Tamazight. ALTRALANG J.; 2024; 6, pp. 227-235.
76. Gajoui, K.E.; Allah, F.A.; Oumsis, M. A Corpus for Amazigh Transcribed to Latin Ocr Systems’Evaluation. Arpn J. Eng. Appl. Sci.; 2006; 13, pp. 8795-8802.
77. El Ouahabi, S.; El Ouahabi, S.; Atounti, M. Comparative Study of Amazigh Speech Recognition Systems Based on Different Toolkits and Approaches. Proceedings of the E3S Web of Conferences; Semarang, Indonesia, 8–9 August 2023; EDP Sciences: Les Ulis, France, 2023; Volume 412, 01064.
78. Klaassen, J.M.; Martin, G.H.; Biler, A.M. The First Translation of the Bible Among Indigenous Islamic Peoples Using a Mediating Approach. 2023; Available online: https://repository.sbts.edu/handle/10392/7100 (accessed on 12 March 2025).
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Details
; Akallouch Mohammed 2
; Fardousse Khalid 1
1 Faculty of Sciences Dhar EL Mahraz, Sidi Mohamed Ben Abdellah University, Fez 30050, Morocco
2 College of Computing, Mohammed VI Polytechnic University, Ben Guerir 43150, Morocco