silero-stress 1.1

Russian Word Stress by Silero

Silero Stress

Silero Stress — pre-trained enterprise-grade automated stress and homograph disambiguation for the Russian language.

Fast start

Dependencies

System requirements to run python examples on x86-64 systems:

• python 3.10+;
• 1G+ RAM;
• A modern CPU with AVX, AVX2, AVX-512 or AMX instruction sets.

Dependencies:

• torch>=1.12.0;

Using pip: pip install silero-stress

from silero_stress import load_accentor
accentor = load_accentor()
sample_sent = "Меня зовут Лева Королев. Я из готов. И я уже готов открыть все ваши замки любой сложности!"
print(accentor(sample_sent))
# Мен+я зов+ут Л+ёва Корол+ёв. +Я +из г+отов. +И +я уж+е гот+ов откр+ыть вс+е в+аши замк+и люб+ой сл+ожности!

Using torch.hub:

import torch
torch.set_num_threads(1)

accentor = torch.hub.load(repo_or_dir='snakers4/silero-stress)
sample_sent = "Меня зовут Лева Королев. Я из готов. И я уже готов открыть все ваши замки любой сложности!"
print(accentor(sample_sent))
# Мен+я зов+ут Л+ёва Корол+ёв. +Я +из г+отов. +И +я уж+е гот+ов откр+ыть вс+е в+аши замк+и люб+ой сл+ожности!

For some configuration tips refer to Configuration wiki page.

You can:

• Use CPU/GPU version;
• Use flags to customize word stress / ё placements;
• Use common accentor and homosolver independently.

Key Features

• Wide coverage

  Silero Stress covers ~4M known Russian words and word forms with 100% accuracy. It also covers ~2K homographs with F1 score of 0.85, per-word accuracy of 91% and total dataset accuracy of 93%.

• Fast

  On 1 CPU thread every ordinary word takes ~0.5ms to process, every 400-char sentence with two homographs (common prod case) takes ~30ms.

• Lightweight and portable

  Total package is about 50MB in size, boasting ~400x compression ratio.

• General and generalizable

  Silero Stress was trained on a large dataset, containing ~4M known words and word forms and ~120M annotated sentences with homographs. It also works with unknown (and invented) words with 60-70% accuracy.

• No Strings Attached

  Published under permissive license (MIT) Silero Stress has zero strings attached - no telemetry, no keys, no registration, no built-in expiration, no keys or vendor lock.

• Minimal and minified

  Code bloat, unnecessary dependencies and libraries are removed. Silero Stress depends only on PyTorch as a neural engine and the Python standard library.

Typical Use Cases

• Academic research
• Text-to-speech applications
• Telephony and call-center automation, voice bots

Links

• Quality Metrics
• Performance Metrics
• Homograph List
• Further reading

Get In Touch

Try our models, create an issue, start a discussion, join our telegram chat, email us, read our news.

Please see our wiki for relevant information and email us directly.

Citations

@misc{Silero Stress,
  author = {Silero Team},
  title = {Silero Stress: pre-trained enterprise-grade automated stress and homograph disambiguation for the Russian language},
  year = {2025},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/snakers4/silero-stress}},
  commit = {insert_some_commit_here},
  email = {[e-mail]}
}

Vocabulary issues

If you see any problems with any particular words, please send your examples as an issue or a discussion.