TS Silero Stress

Russian is a language where the pronunciation lives in the stress marks - and those aren't written down. за́мок is a castle, замо́к is a lock, and a text-to-speech engine that guesses wrong will happily read you the wrong word. TS Silero Stress is a one-input text preprocessor that fixes this: it places stress marks and restores ё (which Russian writing also mostly drops) so the TTS downstream pronounces things right. If you've ever had a Russian voiceover stumble over an ordinary sentence, this is the node that was missing.

How it works

It wraps the silero-stress library, which is actually two independent passes you can toggle:

• use_accentor - a rule-based accentor that handles stress and ё placement for ordinary, non-homograph words.
• use_homosolver - a neural disambiguation model for homographs, the words where stress actually changes meaning (that castle/lock problem). This is the expensive, smart pass.

Each pass has its own put_stress / put_yo pair, so you can, say, restore ё but skip stress marks, or fix only the homographs. stress_single_vowel (on) even marks single-vowel words, which is pedantic but occasionally exactly what an educational worksheet wants. words_to_ignore takes a comma/newline list to skip completely - proper nouns you don't want mangled.

The inputs that matter

• text - your Russian text. That's the one you'll actually type.
• run_device - cpu by default, which is right: this is a small model and CPU is instant. No need to burn GPU on it.
• stress_marker - the output format: unicode (Unicode combining acute accents, the default, which renders cleanly anywhere) or Silero's plus sign notation. Pick unicode unless you're feeding it to something that specifically expects +.
• The put_* / use_* toggles above.

The single output is text - your input with stress marks and ё restored, ready to paste into TS Silero TTS, an SSML block, or a document.

Install

silero-stress is an optional dependency, deliberately outside the core requirements (the Silero packages can conflict with your ComfyUI torch version).

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-timesaver
cd comfyui-timesaver
python -m pip install -r requirements.txt
python -m pip install silero-stress

Weights land in models/silero-stress/ on first use. Skip the install and the node tells you what's missing from the startup log's optional-imports report.

Common issues

• "It left my homographs alone." That's use_homosolver being off, or the homosolver genuinely can't tell from context (it happens with truly ambiguous sentences). Check the toggle first.
• "Output looks like za\\u0301mok garbage in my editor." Those are combining acute accents - the unicode marker. If your downstream tool doesn't render them, switch stress_marker to plus, or use the output only as a pronunciation hint for the TTS (which reads them fine).
• "It marked my brand name." Add it to words_to_ignore.

The honest scope: this is a preprocessor, not a TTS - it produces text, not sound. Chain it in front of TS Silero TTS and you get the full pipeline: correct Russian, pronounced correctly, on your machine, for free.