Feature request - Russian Text Normalization

[freeshard]

🚀 Feature
Support for pronouncing numbers, dates, and currencies in Russian

Motivation
I’ve noticed that when I use the Russian speech synthesis, all digits and numbers are simply skipped or remain silent. It would be great to have this fixed, as currently, I can't hear dates, prices, or any numeric data in the generated audio. This makes it a bit difficult to use the service for reading texts that contain a lot of data or statistics.

Pitch
It would be awesome if you could add a feature that allows the system to recognize and speak numbers in Russian. Ideally, it should:

Voice digits: So that "100" doesn't just result in silence but is spoken as words.

Handle grammar: Russian is tricky with endings, so it would be cool if the numbers sounded natural depending on the sentence (correct cases and genders).

Read dates and money: It would be very helpful if the system could handle dates like 15.05.2026 or prices like 500₽ correctly.

Alternatives
Right now, the only way is to manually rewrite every single number into words (like changing "5" to "пять") before starting the synthesis. It’s quite time-consuming, so having this handled automatically would be a huge lifesaver.

Additional context
Just a few examples of what I'm looking for:

10% -> "десять процентов"

2026 -> "две тысячи двадцать шестой год"

180°C -> "сто восемьдесят градусов"

Thank you for considering this!

[freeshard]

As workarround i update/filter numbers as text locally before send to tts.

_ABBREVIATIONS = [
(re.compile(r'\bмлрд\b', re.IGNORECASE), ' миллиардов '),
(re.compile(r'\bмлн\b', re.IGNORECASE), ' миллионов '),
(re.compile(r'\bкм/с\b', re.IGNORECASE), ' километров в секунду '),
(re.compile(r'\bкм\b', re.IGNORECASE), ' километров '),
(re.compile(r'\bкг.?\b', re.IGNORECASE), ' килограмм '),
(re.compile(r'\bмл.?\b', re.IGNORECASE), ' миллилитров '),
(re.compile(r'\bгр?.?\b', re.IGNORECASE), ' грамм '),
(re.compile(r'\bл.?\b', re.IGNORECASE), ' литров '),
(re.compile(r'\bч.\s?л.?\b', re.IGNORECASE), ' чайных ложек '),
(re.compile(r'\bст.\s?л.?\b', re.IGNORECASE), ' столовых ложек '),
]

def prepare_text_for_tts(text: str) -> str:
"""Cleans up text: removes Markdown, expands abbreviations, numbers → words."""
text = re.sub(r'[*#_`~]', '', text)
text = text.replace('—', '-').replace('–', '-')
text = re.sub(r'(\d+)\s?-\s?(\d+)', r'от \1 до \2', text)

for pattern, replacement in _ABBREVIATIONS:
    text = pattern.sub(replacement, text)

def _num_to_words(match: re.Match) -> str:
    raw = match.group().replace(' ', '')
    try:
        if ',' in raw and raw.endswith(',000'):
            return ' ' + num2words(int(raw.replace(',', '')), lang='ru') + ' '
        if ',' in raw:
            return ' ' + num2words(float(raw.replace(',', '.')), lang='ru') + ' '
        return ' ' + num2words(int(raw), lang='ru') + ' '
    except (ValueError, TypeError):
        return match.group()

text = re.sub(r'\d+(?:\s\d{3})*(?:,\d+)?', _num_to_words, text)
text = re.sub(r'[^\w\s.,!?]', ' ', text)
return ' '.join(text.split())

[MixxxGit]

The task of normalization or verbalization of the Russian language does not have an open solution that provides acceptable quality even at the script level. There are specialized LLMs that supposedly should cope with this task, but in practice they only work well on the data on which they were trained.
The author of Silero knows this problem very well and did not solve it specifically so that users would solve it on their own, as Freeshard did.
The author’s team made attempts to approach this task, but unfortunately failed. see github.com/snakers4/russian_stt_text_normalization

There were other promising attempts, for example, special LLM from sarus72

Unfortunately, it sometimes confuses the declination of numerals and ignores special characters, units of measurement and acronyms.

Currently, the only way is to manually rewrite each number in words (for example, replacing "5" with "five") before starting the synthesis. This is quite a labor-intensive process, so automatically handling this operation would greatly simplify the task.

In fact, any serious LLMs are already coping with this task quite well.
For example DeepSeek Flash

Or even Gemma4

[snakers4]

The author’s team made attempts to approach this task, but unfortunately failed

We made some quick and dirty MVPs with tech available at that time and realized that a proper solution working on CPU (not an LLM, lol) sufficiently quickly would take thousands of man-hours to properly develop and would include several layers of different technologies.

[MixxxGit]

We made some quick and dirty MVPs with tech available at that time and realized that a proper solution working on CPU

It's a shame it didn't work out for you, but I hope your next solution will outshine all the existing ones.
For now, we're left with only what already exists. As it turns out, Python coders recently gave a gift - NickZaitsev/ru-normalizr.
I understand that this is the best approach out of all the existing ones in terms of speed, based on a rule-based architecture, and is specifically tailored for TTS.
But even this Python module has issues and requires further development.

Наш стартап обещает уже в 2026 г. выпустить термостойкий клей, который теряет всего 10% прочности при 180°C.
Наш стартап обещает уже в две тысячи двадцать шестом году выпустить термостойкий клей, который теряет всего десяти процентов прочности при ста восьмидесяти градусах Цельсия.

another sentence

В McDonald's я ходил 10.03.2010 года в 01:59 PM.
В мэкданэлдз я ходил десятого марта две тысячи десятого года года в один, пятьдесят девять пи эм.

or with improved normalization of years

В McDonald's я ходил 10.03.2010 года в 01:59 PM.
В мэкданэлд ' эс я ходил десять целых три сотых. Две тысячи десятого года в один, пятьдесят девять пи эм.