Ukrainian TTS Preprocessing

This repository is an official implementation of the paper: Context-Aware Lexical Stress Prediction and Phonemization for Ukrainian TTS Systems. It contains tools and models for Ukrainian text preprocessing, focusing on lexical stress prediction and phonemization to enhance text-to-speech (TTS) systems, as well as a benchmark dataset for evaluating lexical stress prediction systems for Ukrainian.

We recommend using Python 3.10 or higher for best compatibility. To install all required dependencies, run:

pip install -r requirements.txt

---

Contents

• Ukrainian Lexical Stress Prediction Model
• Ukrainian Phonemizer
• Ukrainian Lexical Stress Benchmark
• Wav2Vec2 with Lexical Stress

---

Ukrainian Lexical Stress Prediction Model

We provide a ByT5-based grapheme-to-phoneme model specialized for predicting lexical stress in Ukrainian words.

Quickstart: Predict Lexical Stress

from src.accentor import UkrainianStressifier

stressifier = UkrainianStressifier()

print(stressifier.apply_stress_marks("Привіт, як у тебе справи?"))

Model Highlights

• Architecture: ByT5 Grapheme-to-Phoneme model
• Training Data: Voice of America corpus, annotated with stress marks by an ASR Wav2Vec2 model

---

Ukrainian Phonemizer

The Ukrainian Phonemizer converts Ukrainian text into phonemes.

Usage Example

from src.phonemizer import UkrainianPhonemizer

phonemizer = UkrainianPhonemizer()

print(phonemizer.phonemize("привіт світе"))

---

Ukrainian Lexical Stress Benchmark

The Ukrainian Lexical Stress Benchmark is a manually annotated dataset created to evaluate lexical stress prediction systems in context.

Dataset location:

lexical_stress_benchmark/data/lexical_stress_dataset.csv

Dataset Format

Each sentence marks stress with a + immediately after the stressed vowel. It contains columns:

• StressedSentence: Sentence with stress annotations
• Source: Origin (wiki, plug, or custom)

Sample Entry

У+ ва+зі стоя+ли кві+ти.,custom

Dataset Statistics

Statistic	Count
Total sentences	1,026
Unique word forms (incl. inflections, derivations)	6,439
Unique words with stress ambiguity (meaning or inflections)	640
Unique words with ≥2 stress forms in dataset	296

Sources

• Wikipedia (300 sentences) — formal encyclopedic style
• Pluperfect GRAC (438 sentences) — fiction, journalism, poetry
• Custom (288 sentences) — manually balanced for ambiguous stress patterns

Evaluation Metrics

• Word-Level Accuracy
• Sentence-Level Accuracy
• Unambiguous Word Accuracy
• Ambiguous Word Accuracy
• Macro-Average F1 (Ambiguous Word Pairs)

Quickstart: Run the Benchmark

from lexical_stress_benchmark.benchmark import evaluate_stressification

def custom_stressify(text):
    """
    Add '+' after the stressed vowel in each stressed word.
    """
    # your implementation here
    return text

accuracies = evaluate_stressification(custom_stressify)
for metric, value in accuracies.items():
    print(f"{metric:40} {value * 100:.2f}%")

---

Wav2Vec2 with Lexical Stress

This model transcribes Ukrainian speech including lexical stress marks directly in the transcription.

• Fine-tuned model on Hugging Face: mouseyy/uk_wav2vec2_with_stress_mark
• Training data: Common Voice corpus annotated with lexical stress from Ukrainian Word Stress and Ukrainian Accentor

---

Citation

@inproceedings{senyk-etal-2025-context,
    title = "Context-Aware Lexical Stress Prediction and Phonemization for {U}krainian {TTS} Systems",
    author = "Senyk, Anastasiia  and
      Lukianchuk, Mykhailo  and
      Robeiko, Valentyna  and
      Paniv, Yurii",
    editor = "Romanyshyn, Mariana",
    booktitle = "Proceedings of the Fourth Ukrainian Natural Language Processing Workshop (UNLP 2025)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria (online)",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.unlp-1.11/",
    doi = "10.18653/v1/2025.unlp-1.11",
    pages = "96--104",
    ISBN = "979-8-89176-269-5",
    abstract = "Text preprocessing is a fundamental component of high-quality speech synthesis. This work presents a novel rule-based phonemizer combined with a sentence-level lexical stress prediction model to improve phonetic accuracy and prosody prediction in the text-to-speech pipelines. We also introduce a new benchmark dataset with annotated stress patterns designed for evaluating lexical stress prediction systems at the sentence level.Experimental results demonstrate that the proposed phonemizer achieves a 1.23{\%} word error rate on a manually constructed pronunciation dataset, while the lexical stress prediction pipeline shows results close to dictionary-based methods, outperforming existing neural network solutions."
}