Kurdish Text-to-Speech
Speech synthesis for Sorani and Badini, built by fine-tuning Meta's MMS TTS. The Sorani model, trained on a curated 10-hour dataset, is the first open Sorani TTS model.
Live demoAI Developer @ Ovanya
Akam Nawzad
Specializing in machine learning and natural language processing.
Selected Work
Publications
-
2026
KuMMLU: A Multitask Language Understanding Benchmark for Central Kurdish
WiNLP 2026 at EMNLP 2026
3,272 multiple-choice questions across 18 subjects, combining native national-exam questions with questions translated from MMLU. I evaluated 32 language models: the best reach about 88% accuracy, and every model does better on translated than on native questions.
-
2025
Fine-tuning XLM-RoBERTa for Named Entity Recognition in Kurmanji Kurdish
WiNLP 2025 at EMNLP 2025
My undergraduate thesis: an XLM-RoBERTa NER system for Kurmanji Kurdish, fine-tuned on 7,919 manually annotated sentences.
Projects
Kurdish Spellchecker
A ByT5 model fine-tuned to correct misspelled Sorani sentences. Working at the byte level makes it robust to the character variation common in Kurdish text.
Live demo
Kurmanji NER Model
The first public Kurmanji NER model and dataset, recognising people, locations and organisations with an F1 score of 0.87.
View model
Handwritten Character Recognition
A CNN written from scratch in NumPy, with manual convolution, im2col and backpropagation. It reaches 71.05% accuracy across 35 handwritten Kurdish characters.
Live demo
Gorrer
A transliterator that converts Kurdish text between the Latin and Arabic scripts.
Visit website
KurdType
A curated collection of open-licensed fonts that support Kurdish, for designers and content creators.
Visit websiteAbout
About me
Machine learning engineer focused on NLP and applied AI.
I work across the full machine-learning lifecycle: collecting and annotating data, training and evaluating models, and deploying them as services. Most of my work is on Kurdish, covering OCR, named entity recognition, spelling, speech and LLM evaluation.
Experience
-
Aug 2025 – present
AI Developer
Ovanya
Building OCR systems and datasets for Sorani Kurdish, including one of the first of their kind. I train and evaluate text-extraction models, write evaluation benchmarks, and deploy inference services with FastAPI.
-
Sept 2024 – Oct 2024
AI Researcher
UKH AI & Innovation Centre
Collected and cleaned a 140K+ article Kurdish news corpus with scalable scraping pipelines, contributed to one of the first Kurdish spell-checking datasets, wrote annotation guidelines, and kept datasets versioned for reproducible research.
Education
-
2021 – 2025
BSc in Software Engineering
University of Kurdistan Hewlêr (UKH)
Graduated in the top 3 of my programme. My thesis on named entity recognition for Kurmanji Kurdish was published at WiNLP 2025.
-
2024
AI Bootcamp
Ovanya
Intensive programme on machine-learning fundamentals and neural networks.
Skills
AI & Machine Learning
- Natural Language Processing
- Deep Learning (CNNs, Transformers)
- Computer Vision & OCR
- Speech Synthesis
- LLM Evaluation & Benchmarks
- RAG & Vector Search
Engineering
- Python
- C++
- SQL
- FastAPI & Flask
- PostgreSQL, MySQL, MongoDB
- Docker
Libraries & Tools
- PyTorch
- TensorFlow
- Hugging Face
- scikit-learn
- NumPy, Pandas, NLTK
- Git, GitHub, GitLab
Contact
Working on Kurdish NLP, or need a hand with a machine-learning project? I'd be glad to hear from you.
akam.nawzad@ukh.edu.krd