Papers

A historical Kurdish text-to-speech study comparing concatenative synthesis systems based on allophone, syllable, and diphone units. The retrieved primary bibliographic metadata verifies the publication and comparison task, while detailed corpus characteristics and exact numerical evaluation results …

Development of a Universal Dependencies treebank for Kurmanji Kurdish using morphological analysis and manual annotation, providing syntactic training and evaluation data for Kurdish language processing.

A contextual Sorani Kurdish spelling-correction study using a trigram language model with Stupid Backoff and edit-distance candidates. The authors report 88.54% overall accuracy and 43.33% F1, with substantially different performance across the educational-text evaluation subsets.

This 2026 Research Square preprint presents a modified Vision Transformer approach for handwritten Sorani Kurdish line recognition and introduces DASNUS. The dataset contains 11,475 lines from 867 writers with writer-separated training, validation and test sets. The authors report 3.47% CER and …

A Kurdish tokenization system using lexical resources and morphological analysis to address token boundaries, compounds and orthographic variation in Sorani and Kurmanji.

This dataset paper introduces K-PIT and K-ZHMARA for isolated Central Kurdish handwriting recognition. K-PIT contains 245,000 character images across 35 classes and K-ZHMARA contains 70,000 digit images across 10 classes, for 315,000 images in total. The publication presents the resources as …

A Sorani Kurdish speech synthesis study adapting F5-TTS with a specialized single-speaker collection derived from broadcast material. The paper reports subjective and recognition-based evaluation; the underlying dataset, Kurdish checkpoint, and code were not independently verified as public …

This Data in Brief article describes a collection of isolated handwritten Central Kurdish characters. The paper reports 40,940 images from 390 writers and tables containing 35 labels. A later published letter reported 40,826 images in examined earlier releases and identified inconsistencies in the …

A corpus-creation study for severely under-resourced Southern Kurdish varieties and Laki using web extraction, radio material and digitized fieldwork.

Develops an extractive automatic text summarization approach for long-form Sorani Kurdish research documents. The study uses a corpus of 231 scientific papers from four academic departments, applies Sentence Weighting and TF-IDF in two experiments with and without conclusion sections, and evaluates …

Introduces a parallel news corpus for English, Turkish and Northern Kurdish (Kurmanji) and evaluates bilingual and multilingual neural machine translation using Nematus. The work provides an early verified neural machine translation resource involving Kurdish.

Develops end-to-end Transformer-based automatic speech recognition for Central Kurdish using a newly constructed 100-hour speech corpus. The study fine-tunes wav2vec 2.0 XLS-R models and combines them with n-gram language models and a Kurdish tokenizer. The best reported configuration achieved 10.0% …

A Sorani Kurdish lemmatization and word-level spelling-correction study. The Rênûs spell checker combines character n-grams, corpus frequencies, lexicon lookup and weighted edit distance. The authors report a best first-ranked correction accuracy of 96.4% with a lexicon and 87% without a lexicon.

Develops a Central Kurdish automatic speech recognition system using deep learning and transfer learning. The approach combines MFCC speech features with LSTM networks and Connectionist Temporal Classification and uses the AsoSoft speech database.

A Central Kurdish text-to-speech study presenting three separately fine-tuned single-speaker F5-TTS models trained on two audiobook-derived resources and reused Gigant recordings. The study evaluates synthesized speech with human listeners and applies synthetic speech to downstream speech …

An end-to-end Central Kurdish text-to-speech study using latent speech representations, text-audio alignment, adversarial waveform training, and duration prediction. The system is evaluated on the Gigant-KTTS corpus. Reported quality and speed measurements belong to the cited experiment and are not …

Development of an electronic lexical resource for Southern Kurdish from the Başûr dictionary using OntoLex-Lemon, with exploration of lexical alignment with Sorani Kurdish.

A Central Kurdish speech synthesis study combining a CNN-Transformer encoder, Tacotron 2 autoregressive decoding, and WaveGlow. The experiment reuses Gigant-KTTS recordings and evaluates synthesized speech with native listeners and word error rate measurements.

Development of the Kurdish Textbooks Corpus, a normalized Sorani corpus built from K-12 textbooks and organized across twelve educational subject areas.

A Sorani Kurdish speech synthesis study using transfer learning from an English Tacotron 2 model with a pretrained HiFi-GAN vocoder. The experiment uses a single female speaker corpus of approximately ten hours and evaluates synthesized speech with native listeners.

Investigates end-to-end automatic speech recognition for Northern Kurdish (Kurmanji) by fine-tuning OpenAI Whisper v3 using multiple adaptation strategies. The work uses the Northern Kurdish portion of Mozilla Common Voice 18.0 and reports a best WER of 10.5% and CER of 5.7%.

This study adapts an Arabic Tesseract LSTM model to printed Sorani Kurdish using limited Kurdish ground truth and synthetic font-based training material. The authors report 95.45% average character accuracy on an evaluation set of 11 court-document images. A separate Tesseract evaluation reports …

A qualitative English-Kurdish study of lexical gaps in five theoretical-physics and popular-science books and their Kurdish translations, approximately 520,000 words as reported by the authors. It examines scientific words, eponyms and abbreviated forms, and discusses responses including triplets, …

Extends the multilingual FLEURS benchmark to Northern Kurdish with 5,162 validated utterances totaling 18 hours and 24 minutes from 31 native speakers. The resource supports ASR and speech-translation evaluation; the reported Whisper v3-large ASR baseline achieved 28.11% WER and 9.84% CER.

A dataset paper describing Gigant-KTTS, a Central Kurdish speech corpus constructed specifically for text-to-speech research. The corpus contains 6,078 recorded sentences across twelve subject categories from one male speaker.

Development of a Sorani Kurdish Hunspell resource combining a morphosyntactically annotated lexicon with morphological rules for spelling correction, stemming and morphological analysis.

A rule-based Hunspell resource for Sorani Kurdish spelling suggestions and morphological analysis. The 2021 paper reports a 23,223-entry lexicon and 4,293 rules. Its reported evaluation should be interpreted as the performance of the evaluated 0.1.0 version, not current releases.

Jira is a large-vocabulary speech recognition system for Central Kurdish (Sorani), developed together with a purpose-designed speech corpus and pronunciation lexicon. The work uses corpus design based on diphone distribution, acoustic modeling, and n-gram language modeling, reporting a word error …

This Data in Brief article introduces a scene-text dataset covering Kurdish Sorani and Badini in Arabic-based script alongside Arabic and English. It reports 1,420 natural scene images, 19,872 real word samples and 20,000 synthetic instances across the multilingual collection. The original KSTRV1 …

Presents an early experimental speech-to-text resource for Sorani Kurdish using CMUSphinx. The accompanying BD-4SK-ASR dataset contains 2,000 utterances derived from vocabulary and sentences in primary-school educational material.

A Central Kurdish end-to-end speech synthesis study comparing Tacotron 2 and WaveGlow configurations, including transfer learning from English and training from scratch. The work uses a single-speaker female Kurdish speech corpus associated with the AsoSoft corpus family.

A Kurmanji Kurdish study combining morphological lemmatization, character bigrams and Jaccard similarity for word-level spelling correction. The authors report 90.77% correction accuracy. The paper's test-sample descriptions are internally inconsistent, so the result should be presented with that …

This Data in Brief article presents an EMNIST-like handwritten character dataset for Sorani Kurdish. The resource contains 58 character-shape categories and the repository reports 406,000 images. The publication verifies Sorani coverage; broader claims of Latin Kurmanji coverage in repository …

Introduces KurdSum, a Kurdish news summarization benchmark containing more than 40,000 texts paired with human-written summaries. The study evaluates four extractive approaches (LexRank, TextRank, Oracle, and Lead-3) and three abstractive approaches (Pointer-Generator, Sequence-to-Sequence, and …

Development of CORDI, a variety-sensitive Central Kurdish corpus containing transcripts and regional metadata from movies and television across six urban varieties.

A study of language conventions in selected material from five Kurdish scientific books. Using Lambert and van Gorp's framework, the authors examine scientific-language conventions, specialist vocabulary, English loans, nominalization and variation in language choices. Terminology is one component …

Develops manually validated parallel news data for Sorani, Kurmanji and English using multilingual news websites. The associated corpus contains Sorani-Kurmanji, English-Kurmanji and English-Sorani aligned text.

Introduces LR-Sum, an openly licensed multilingual summarization dataset containing human-written summaries for 40 languages. Kurdish is represented as separate Kurmanji and Sorani varieties. The paper establishes extractive and abstractive baselines, including mT5-based summarization, for …

This 2024 arXiv preprint adapts Tesseract 5 for historical printed Sorani Kurdish publications using transcribed lines from pre-1950 material. The abstract reports 1,233 line images with transcriptions, while the currently inspected partial release contains 1,203 TIFF images and 1,203 text files. …

A descriptive study of medical terminology translation using 150 English patient information leaflets and their Kurdish translations. The authors identify nine translation procedures, reporting borrowing as the most frequent and modulation as the least frequent. The findings describe the studied …

A conceptual study of terminology-dictionary construction using terminology from novel criticism as a case study. It emphasizes concept analysis and differentiating meanings when selecting terminology rather than simply collecting, borrowing or translating everyday words. Its scope is literary …

Presents community-created parallel corpora for low-resource Middle Eastern language varieties and machine translation experiments using NLLB-600M. Kurdish-related resources include Southern Kurdish, Laki, Hawrami and Zazaki data as identified by the publication.

Presents the No Language Left Behind multilingual machine translation system. The released model family explicitly supports Central Kurdish in Arabic script and Northern Kurdish in Latin script among approximately 200 supported languages.

A multilingual speech technology study associated with Meta's Massively Multilingual Speech project. Its Kurdish relevance includes released Northern Kurdish/Kurmanji VITS synthesis checkpoints with Latin, Arabic, and Cyrillic script identifiers. The official coverage information does not establish …

A methodological study identifying limitations of the existing KurdNet prototype and proposing a semi-automatic approach for further development of the Kurdish lexical-semantic database.

An empirical comparison of Sorani and Kurmanji based on the Pewan news corpus, examining linguistic, lexical and orthographic characteristics across the two Kurdish varieties.

Introduces speech-recognition resources and an evaluation benchmark for Southern Kurdish, including approximately 30 hours of validated read speech. Whisper-turbo and wav2vec-BERT CTC baselines are evaluated, with the wav2vec-BERT system reporting 4.09% CER and 24.26% WER on the benchmark.

A Central Kurdish spelling-correction method using a compact pattern dictionary and recursive Levenshtein-distance matching. The authors report the intended correction ranked first in 79.93% of synthetic word-level tests, within the top three in 93.30%, and within the top five in 97.01%.

A qualitative discussion of scientific terminology in Kurdish, including terminology development, word formation, and the distinction between linguistic and terminological meaning. The article discusses obstacles to developing scientific terminology and argues for coordinated work by specialists and …

A descriptive lexicological study examining specialist terminology, technical coinages, foreign terminology, and lexicalization in Kurdish. The study explains how specialist terms can enter and enrich the general lexicon. It does not establish a standardized terminology database.

Presents a multilingual medical and crisis-domain translation benchmark containing professionally translated and quality-controlled material. The release includes separate English-Sorani and English-Kurmanji evaluation sets.

Development and analysis of a large Central Kurdish text corpus collected from websites, books and magazines, with Kurdish-specific normalization and TEI-based organization.

Development of KurdNet, a Kurdish lexical-semantic resource aligned with Princeton WordNet concepts. The implemented prototype focuses on Sorani Kurdish and uses bilingual lexical material and corpus examples.

A study of electronic lexicography for Kurdish that converts lexicographic material into structured machine-readable resources using OntoLex-Lemon and RDF, covering Sorani, Kurmanji and Hawrami lexical material.

Investigates Sorani Kurdish-English machine translation using OpenNMT LSTM encoder-decoder models and several tokenization strategies. The experiments use Tanzil, TED and translated KurdNet material and demonstrate strong domain sensitivity.

Evaluates unadapted multilingual automatic speech recognition systems on a newly constructed Garrusi Kurdish evaluation set. The study uses staged text normalization to enable more consistent CER and WER comparison across orthographic variation.