README.md

Belarusian NLP and Speech Processing resources

This repository contains links to Belarusian Natural Language and Speech Processing resources and datasets.

It is inspired by similar project with Ukrainian Speech Processing resources: egorsmkv/speech-recognition-uk

🧍‍♀️🧍 Communities and platforms:

• say.by - web page, telegram
• corpus.by
• ssrlab.by
• bnkorpus.info
• Belarus organization on github
• nlproc.by community on github

💬➡️📘 Speech-to-Text

💬➡️📘 - 📚 Datasets

• Common Voice. Speech recognition dataset
• Dataset from knihi.com. TODO: what is the type of dataset?
• google/fleurs
• ssrlab: TODO. Speech recognition dataset

💬➡️📘 - 🚀 Models

• wav2vec2 trained on Common Voice 8 + kenlm language model trained on Common Voice 8:

  • Model: ales/wav2vec2-cv-be
  • Demo: ales/wav2vec2-cv-be-lm
  • Code: navalnica/wav2vec2-belarusian

• whisper:

  • original openai/whisper models
  • Whisper models fine-tuned on Belarusian Common Voice 11 dataset:
    • Whisper Small:
      • Model: ales/whisper-small-belarusian
      • test WER on CommonVoice11: 6.79
      • Demo: ales/whisper-small-belarusian-demo
      • Code: navalnica/whisper-finetuning-be
    • Whisper Base:
      • Model: ales/whisper-base-belarusian
      • Code: navalnica/whisper-finetuning-be

• Nvidia NeMo models:

  • nvidia/stt_be_conformer_ctc_large
    • [huggingface self-reported metric] test WER on CommonVoice10: 4.8
  • nvidia/stt_be_conformer_transducer_large
    • [huggingface self-reported metric] test WER on CommonVoice10: 3.8
  • nvidia/stt_be_fastconformer_hybrid_large_pc
    • [huggingface self-reported metric] test WER on CommonVoice12: 2.72
    • [huggingface self-reported metric] test WER P&C CommonVoice12: 3.87

• ESPnet:

  • espnet/belarusian_commonvoice_blstm

💬➡️📘 - 🥇 Benchmarks

Model comparisons grouped by dataset. TODO

---

📘➡️💬 Text-to-Speech

📘➡️💬 - 🚀 Models

CoquiAI implementations

• official CoquiAI contains recipe for Belarusian
• jhlfrfufyfn/bel-tts. GlowTTS + HifiGan
  • Code
  • Model
  • Demo on HuggingFace
  • Demo on a custom web-page. The source code for the demo page: here
• alex73/belarusian-tts. CoquiAI implementation by Yurii Paniv (@robinhad).
  Original repo & models were deleted - only fork is available now

---

📝 NLP

📝 - 📚 Datasets

• BelarusianGLUE - dataset, paper
• oscar
• mc4
• poritski/YABC - Эксперыментальны корпус беларускай мовы, ЭКБМ
• Belarus/GrammarDB - Grammar Database of Belarusian language
• tsimafeip/Translator - Dataset with russian-belarusian translation pairs
• Universal dependencies dataset:
  • Page
  • GitHub Repository
• Tatoeba Belarusian sentences

📝 - 🚀 Models

POS-tagging

• KoichiYasuoka/roberta-small-belarusian-upos
• stanfordnlp/stanza-be
• poritski/YABC_Tagger. Rule-based POS-tagger and lemmatizer.
  Written in Perl. Uses poritski/YABC as a Grammar base (?)
• volchek/beltagger. An improved version of poritski/YABC_Tagger rule-based POS-tagger and lemmatizer.
  Cross-platform, written in C++.
  Known issues:
  • requires input data to be incoded in Windows-1251, does not support UTF-8;
  • tagset is not fully-compatible with BNKorpus's tagset and grammar base
  • grammar base used is not full enough. Belarus/GrammarDB is a better paradigms source but is not incorporated yet
  • suffix table calculation script is not ported from Perl to C++
  • code uses Boost libarary

Other

• pkasila/bel-sklony - web page with Belarusian nouns declension. Demo: sklony.pkasila.net

Masked Language Modeling

• KoichiYasuoka/roberta-small-belarusian