Belarusian NLP and Speech Processing resources

A curated list of Belarusian Natural Language and Speech Processing resources, datasets, and models.

Inspired by egorsmkv/speech-recognition-uk.

Table of Contents

• 🤝 Contributing
• 🏡 Communities and platforms
• 📚 Datasets
• 🎤 Speech-to-Text Models
• 🤖 Text-to-Speech Models
• 📝 NLP Models and Tools

🤝 Contributing

Contributions are very welcome! Please open a pull request to add missing resources, update entries, or fix any of the TODOs below. Help keep this list current.

🏡 Communities and platforms

• say.by — Telegram
• corpus.by
• ssrlab.by
• bnkorpus.info
• Belarus — GitHub organization
• nlproc.by — GitHub community

📚 Datasets

Name	Type	Notes
Common Voice	STT	Mozilla multilingual speech corpus
google/fleurs	STT
knihi.com — Корпус беларускага маўлення	STT	Corpus for neural network training. TODO: confirm dataset type
ssrlab	STT	TODO
maaxap/BelarusianGLUE	NLP	Paper
oscar-corpus/oscar	NLP
allenai/c4	NLP
poritski/YABC	NLP	Эксперыментальны корпус беларускай мовы (ЭКБМ)
Belarus/GrammarDB	NLP	Grammar Database of Belarusian language
tsimafeip/Translator	NLP	Russian-Belarusian translation pairs
UniversalDependencies/UD_Belarusian-HSE	NLP	Universal Dependencies treebank. Project page
Tatoeba — Belarusian	NLP	Belarusian sentences in Tatoeba

🎤 Speech-to-Text Models

Name	Metrics	Notes
ales/wav2vec2-cv-be	WER 12.4 on test split of CV8	wav2vec2 + kenlm language model, trained on Common Voice 8. Demo, Code
ales/whisper-small-belarusian	WER 6.79 on test split of CV11	Whisper Small fine-tuned on Common Voice 11. Demo, Code
ales/whisper-base-belarusian	WER 12.207 on validation split of CV11	Whisper Base fine-tuned on Common Voice 11. Code
openai/whisper	-	Original Whisper models from OpenAI
nvidia/stt_be_conformer_ctc_large	WER 4.8 on test split of CV10	NVIDIA NeMo Conformer-CTC
nvidia/stt_be_conformer_transducer_large	WER 3.8 on test split of CV10	NVIDIA NeMo Conformer-Transducer
nvidia/stt_be_fastconformer_hybrid_large_pc	WER 2.72 on test split of CV12 WER P&C 3.87 on test split of CV12	NVIDIA NeMo FastConformer Hybrid
espnet/belarusian_commonvoice_blstm	-	ESPnet BLSTM trained on Common Voice

🤖 Text-to-Speech Models

Name	Metrics	Notes
coqui-ai/TTS	-	Official CoquiAI Belarusian recipe
jhlfrfufyfn/bel-tts	-	GlowTTS + HifiGAN. Model, HF demo, web server source code
alex73/belarusian-tts	-	CoquiAI implementation by Yurii Paniv (@robinhad). Original repo and models deleted — only fork remains

📝 NLP Models and Tools

Name	Topic	Notes
alesdrobysh/belmorph	Morphological analysis	Fast TypeScipt morphological analyzer, inflection, and lexeme generator. Demo, a really beautiful one!
pkasila/bel-sklony	Declension (вызначэнне меснага склону)	Belarusian nouns declension (месны склон only) web page. Demo
poritski/YABC_Tagger	POS-tagging, Lemmatization	Rule-based. Perl. Uses poritski/YABC as grammar base
volchek/beltagger	POS-tagging, Lemmatization	Improved C++ port of poritski/YABC_Tagger. Cross-platform. Known issues: requires Windows-1251 input (no UTF-8), tagset not fully compatible with BNKorpus, incomplete grammar base (Belarus/GrammarDB not yet incorporated), suffix table calculation not ported from Perl, depends on Boost
stanfordnlp/stanza-be	POS-tagging	Stanza framework
KoichiYasuoka/roberta-small-belarusian-upos	POS-tagging	roberta model
KoichiYasuoka/roberta-small-belarusian	Masked Language Modeling	roberta model