Up to $50,000 Funding
Grant support for language data collection for low-resource European languages.
LINGUA Open CallMicrosoft AI for Good
Building Ukrainian Construction Language for AI
BUDOVA: Building Ukrainian Domain-Specific, Open Voice & Text Archives — the first open Ukrainian construction-domain language dataset for speech and text to advance AI technologies in Ukraine.
01
100+
hours of speech (target)
02
100M+
text tokens (target)
03
524B
USD reconstruction needs (RDNA4 estimate, 2025)
04
CC BY 4.0
license
Animated promo, 23 seconds, music only, no narration. On screen: built in Ukrainian; AI speaks English; zero construction datasets. Step 1: speech is recorded on site. Step 2: speech and DBN building codes are cut into segments. Step 3: NER annotation in three passes with nine labels, 2,492 segments. Step 4: export as JSONL for AI. Target text corpus: 100M+ tokens. Building Ukrainian construction language for AI. BUDOVA, budov.org, supported by Microsoft AI for Good Lab.
Initiative
LINGUA Open Call
A Microsoft AI for Good Lab initiative under EU Digital Unlock, supporting digital inclusion for low-resource European languages. The open call selected 11 projects covering 16 languages and dialects across 10 countries.
Microsoft
Azure Compute Resources
Cloud computing credits for up to 2 years for data processing and validation.
Technical Support
Research collaboration with Microsoft AI for Good Lab.
Open Models
Integration with Apertus, EuroLLM, SmolLM3 and other multilingual models.
Challenge
Why This Is Critical
Despite 30–46 million speakers, Ukrainian remains critically underrepresented in AI technologies — especially in specialized domains.
<0.6%
of web content in Ukrainian
Digital Exclusion
No technical language datasets, speech recognition systems for construction sites, or AI tools supporting Ukrainian building codes.
3
dialect groups, 15+ subdialects
Dialect Diversity
Northern, Southwestern (Volhynian, Podolian, Upper Dniester, Upper Sian, Pokuttia–Bukovyna, Hutsul, Boyko, Transcarpathian, Lemko), and Southeastern dialect groups remain undocumented in technical contexts — AI systems risk failure for non-standard speakers.
15–25%
performance gap
Morphological Complexity
Cyrillic script, 7 cases, 3 genders, moveable stress — unique tokenization challenges with a performance gap vs. English.
What We Build
Open Resources for Construction
Our goal: the first comprehensive Ukrainian technical language dataset with dialectal coverage across all major regions of Ukraine.
Speech
Speech Dataset
Target: 100+ hours of annotated construction speech with transcriptions, covering Northern, Southwestern, and Southeastern Ukrainian dialects. Read-aloud collection opened in October 2026.
<25%WER · target
>0.70Expert kappa · target
Text
Text Corpus
Target: 100M+ tokens of raw construction text — building codes, legislation, estimating norms, open-access journals. The NER-annotated subset already holds 2,927 segments from nine building codes (ДБН).
>0.85NER F1 · target
5–10Kterms · target
Infrastructure
Technical Infrastructure
Baseline ASR and NER models, data processing pipelines, documentation per Datasheets for Datasets standards. To be published on Hugging Face & Zenodo.
JSONLtext format
DOIZenodo · planned
Roadmap
24 Months, 5 Phases
A phased plan with quarterly releases for community feedback. The first public data release is v1.0 (Q4 2026).
01 / 05Phase 1
Infrastructure & Preparation
Months 1–3 · Jan – Mar 2026Infrastructure setup, GDPR protocols, participant recruitment.
02 / 05Phase 2
Data Collection
Months 4–15 · Apr 2026 – Mar 2027Plan: 25 h of speech + 25M tokens per quarter toward the 100+ h and 100M+ token targets. Status, October 2026: 9 building codes annotated (2,927 segments); read-aloud speech collection has started.
03 / 05Phase 3
Annotation & Validation
Months 6–18 · Parallel with collectionTwo-stage validation: crowdsourced transcription, then expert review. Target: human expert agreement κ > 0.70; text annotation passes have a separate agreement threshold, κ ≥ 0.8.
04 / 05Phase 4
Baseline Model Development
Months 12–21 · Dec 2026 – Sep 2027Training ASR and NER models on collected data. Target: WER <25% for construction terminology.
05 / 05Phase 5
Release & Transfer
Months 21–24 · Sep – Dec 2027Final dataset release on Hugging Face and Zenodo with DOIs. Documentation, sustainability transition.
Datasets
Open Research Resources
Large-scale datasets and benchmarks for training, evaluating, and testing Ukrainian construction-domain NLP models.
Speech Dataset
100+hours of speech (target)
Target: transcribed construction speech from all major Ukrainian dialect groups. Today contributors read short construction texts aloud right in the browser.
Hugging Face · planned for a later releaseText Corpus
100M+text tokens (target)
Raw construction text — building codes, legislation, estimating norms, open-access journals. The NER-annotated subset holds 2,927 segments from nine building codes (ДБН).
Hugging Face · first sample with v1.0 (Q4 2026)Construction Lexicon
5–10Kdomain terms (target)
Domain-specific terminology covering building materials, techniques, safety standards, and regulatory language across construction sub-domains.
Hugging Face · planned for a later releaseLanguageUkrainian (uk-UA) · Northern, Southwestern & Southeastern dialects
License (planned)CC BY 4.0 (data) · Apache 2.0 (models) · MIT (code)
HostingHugging Face · Zenodo (DOI) — from v1.0 (Q4 2026) · GitHub
PrivacyGDPR · Voice anonymization & PII removal — Roadmap
Examples
Live examples from the corpus.
Real segments from the corpus — Ukrainian building codes annotated with nine entity labels and added to the platform. Filled highlight: the example’s focus label; outlined: the other annotated entities.
«Теплоізоляційний шар горищного перекриття для дахів слід передбачати з матеріалів груп горючості НГ або Г1 згідно з ДБН В.1.1-7. Товщину шару необхідно приймати відповідно до теплотехнічного розрахунку згідно з ДБН В.2.6-31.»
- Structure
- Теплоізоляційний шар · горищного перекриття · дахів
- Material
- матеріалів
- Process
- теплотехнічного розрахунку
- Measurement
- НГ · Г1
- Property
- груп горючості
- Regulation
- ДБН В.1.1-7 · ДБН В.2.6-31
- Source
- ДБН В.2.6-220:2017 · Roofs of buildings and structures
- In English
- The thermal-insulation layer of the attic floor for roofs shall be made of materials of combustibility group NG or G1 in accordance with DBN V.1.1-7. The layer thickness shall follow the thermal-engineering calculation in accordance with DBN V.2.6-31.
Open Resources
Full Transparency. No Restrictions.
All outputs — data, models, code — will be released under open licenses.
Datasets
Speech and text corpora with annotations and metadata, plus a construction lexicon (target: 5–10K terms).
Baseline Models
Roadmap: fine-tuned ASR (Whisper / Wav2Vec2) and NER models for the construction domain with reproducible scripts — after v1.0.
Infrastructure
Roadmap: processing pipelines, anonymization utilities and validation workflows — for replication by other low-resource projects.
Ready to Get Started?
The datasets will be published on Hugging Face and Zenodo — the first text-corpus sample with v1.0 (Q4 2026). Until then, preview the data format and live samples on the dataset page.
Preview the datasetsHow to Cite
If you use BUDOVA resources in your research, please cite:
@misc{budova2026,
title = {{BUDOVA}: Building {Ukrainian} Domain-Specific, Open Voice \& Text Archives},
author = {Dolhopolov, Serhii},
year = {2026},
howpublished = {\url{https://budov.org}},
organization = {KNUCA AI Lab, Kyiv National University of Construction and Architecture},
note = {Pre-release; first public dataset release (Hugging Face and Zenodo) planned for Q4 2026}
}Deliverables
Project Outputs
Concrete research outputs and open resources produced throughout the project lifecycle.
01 / 04
3datasets
Dataset Releases
Versioned speech and text datasets, to be published on Hugging Face and Zenodo with full documentation and DOIs.
25%ETA Q4 2026
- Infrastructure & GDPR protocols
- First 25h speech collected
- Text corpus v0.1
02 / 04
2+models
Baseline Models
Fine-tuned ASR (Whisper / Wav2Vec2) and NER models for Ukrainian construction domain with reproducible scripts.
5%ETA Q2 2027
- Model architecture selected
- Training pipeline setup
- Benchmark evaluation
03 / 04
3+papers
Scientific Papers
Peer-reviewed publications at NLP and AI conferences documenting methodology, benchmarks, and findings.
10%ETA Q3 2027
- Literature review completed
- Methodology paper draft
- Conference submission
04 / 04
100%open-source
Open-Source Tools
Processing pipelines, anonymization utilities, validation workflows, and documentation per Datasheets for Datasets standards (the first datasheet ships with v1.0).
15%ETA Q4 2027
- Repository structure created
- Anonymization pipeline
- Validation workflows
Research Group
Project Team
An interdisciplinary team with expertise in AI, NLP, construction, cybersecurity, and data management.
01 / 01Project Team Lead
Serhii Dolhopolov
KNUCA AI Lab · Ph.D. Student in Computer Science
AI researcher and entrepreneur specializing in Natural Language Processing. Founder of KernelGlide — AI solutions for 10+ clients in construction. Principal Investigator on state-funded research (UAH 2.9M) in multimodal content analysis. Author of the textbook "Modeling AI Tasks" (546 pp.).
Top Contributors
- The contributor ranking loads here
Host institution
knuba.edu.uaKyiv National University of Construction and Architecture
Founded in 1930 · nearly a century of leadership in construction education & research
