This is a replacement for the folder https://huggingface.co/datasets/styletts2-community/multilingual-pl-bert/tree/main/be. I haven't changed `input_ids`, only `phonemes`. ~95% of all transcriptions were amended where appropriate; the remaining ~5% were copied (with minor changes in notation) from the original version of the dataset.

A few caveats:
- My amended transcriptions are based on the outputs of two existing Belarusian G2P tools, [corpus.by](https://corpus.by/TranscriptionGenerator/?lang=en) and [bnkorpus.info](https://bnkorpus.info/fanetyka.html). They use slightly different IPA character sets, so I combined the outputs by applying the same formatting to both.
- I follow these conventions to represent Belarusian sounds in IPA: [о] is `ɔ`, [э] is `ɛ` (no matter if the preceding consonant is velar or palatal); velar [л] is `ɫ`, palatal [л'] is `lʲ`, [ч] is `t͡ʂ`, [ц] is `t͡s`, plosive [ґ] is `ɡ`.
- Stress marks precede stressed syllables, i.e. there may be one or more consonants between the mark and the stressed vowel.
- Punctuation is (mostly) preserved in the transcriptions, and `[UNK]` token is represented by empty string.
- Arabic and roman numbers are transcribed as their corresponding Belarusian numerals. There may be errors in gender / number / case, as the contexts are very diverse. Similarly, some abbreviations (several dozen most frequent ones) are transcribed as their corresponding full words or phrases.
- Words in Russian and English are transcribed using espeak-ng. No separate processing was applied to words in other languages that incidentally occur in Belarusian texts.
- Among the ~5% hard instances which I wasn't able to refine, most are proper names, alphanumeric strings, URLs, etc. Also, I didn't have enough time to disambiguate variably-stressed items, like *му́зыка* 'music' vs. *музы́ка* 'musician'; their transcriptions are also copied from the original dataset.