Support Unicode stress marks (U+0301) for Russian text

[stilletto]

Problem

When generating Russian speech, OmniVoice sometimes places word stress incorrectly. Russian has many homographs where stress position changes meaning:

• за́мок (castle) vs замо́к (lock)
• му́ка (suffering) vs мука́ (flour)
• ду́хи (spirits) vs духи́ (perfume)

Currently there is no way to control stress placement in Russian text.

What I tried

Method	Result
Unicode combining acute accent (U+0301): за́мок	Ignored by model
Plus sign after vowel: за+мок	Ignored or read as literal "+"
ARPAbet injection [Z AA1 M AH0 K] for single words	Works! But gives English accent
Full text in ARPAbet	Sounds foreign, not Russian
SSML <phoneme> tags	Ignored

The only working approach is partial ARPAbet injection (specific homograph words only, rest in Cyrillic), but it introduces a slight English accent on those words.

Proposed solution

Support the Unicode combining acute accent (U+0301, ◌́) placed after the stressed vowel in Russian text:

"Старый за́мок стоял на холме́. Замо́к на двери́ был сло́ман."

This is the standard linguistic way to mark Russian stress and is already used by:

• Russian dictionaries
• ruaccent library
• Wiktionary Russian entries

Why this matters

Russian TTS quality depends heavily on correct stress. With 600+ language support, OmniVoice is uniquely positioned to handle this well. A simple stress mark would make Russian output significantly more natural, especially for:

• Audiobook narration
• Game NPC dialogue
• Virtual assistants

Environment

• OmniVoice 0.1.2
• PyTorch 2.8.0, CUDA
• Using voice cloning with Russian reference audio

[babaevlsdd]

I can confirm that stress marks (diacritics) actually work on my side.

When I explicitly add stress (e.g., “за́мок” vs “замо́к”), the model produces the correct pronunciation consistently. So it doesn’t seem like the model is completely ignoring them.

[gecko984]

There are at least two earlier issues about the same thing: #37 and #52

As far as I can tell, the nature of how the model does training and inference, there is no way to hardcode any such behaviours. All we can do is create a dataset with pairs (text with stress marks, audio), and finetune the model with it

[gecko984]

In my experience, the stress mark U+0301 already works, but only to some degree, ruining the pronounciation of some words

[stilletto]

There are at least two earlier issues about the same thing: #37 and #52

As far as I can tell, the nature of how the model does training and inference, there is no way to hardcode any such behaviours. All we can do is create a dataset with pairs (text with stress marks, audio), and finetune the model with it

I am preparing such dataset with stress marks. Almost done with Ruslan dataset. 80% done. But it is only one voice and not sure how many osmographs it is cover.

[stilletto]

In my experience, the stress mark U+0301 already works, but only to some degree, ruining the pronounciation of some words

May be I do something different but in my samples the model absolutely ignore this stress mark.

[SafeCodee]

Maybe solution is to convert input text using lib like https://github.com/snakers4/silero-stress or some similar
chatterbox is using russian-text-stresser sure for ru language which is not up-to-date and there is an isssue about it
But it's not solution if it doesn't support U+0301

[zhu-han]

There seem to be many duplicate issues regarding stress control in Russian. I will close most of them and keep only the most recent one: #129