UNLP · 2025 · Workshop
Senyk et al. (Ukrainian Catholic University; Taras Shevchenko National University of Kyiv) · → Paper · Demo: ? · Code: ✓
A rule-based phonemizer and context-aware ByT5-based lexical stress prediction model for Ukrainian TTS preprocessing, together with the first publicly released benchmark dataset for Ukrainian lexical stress evaluation at the sentence level.
Problem
Ukrainian TTS systems are hampered by weak text preprocessing. Ukrainian poses two compounding difficulties: a non-deterministic stress system in which many words have multiple valid stress placements depending on morphological or syntactic context (heteronyms, inflectional stress shifts, free variants), and rich phonology requiring faithful grapheme-to-phoneme conversion. Existing phonemizers for Ukrainian either overgeneralise assimilation rules, require user intervention at morpheme boundaries, or apply naïve letter-to-phoneme mapping that ignores phonological phenomena entirely. Stress prediction tools are similarly constrained: word-level neural models fail to resolve contextual ambiguity, and dictionary-based pipelines cannot handle out-of-vocabulary words. No publicly available benchmark existed for measuring progress on sentence-level stress prediction.
Method
The paper introduces two complementary preprocessing components and a benchmark dataset for Ukrainian TTS.
The phonemizer is a rule-based grapheme-to-phoneme system implemented in Python using ordered regular expressions. Rules are applied sequentially to the full input string without tokenization, covering: initial grapheme-to-symbol mapping (including multi-character units such as дз, дж), consonant cluster reduction, voicing assimilation, sibilant assimilation, palatalization assimilation, and allophonic variation. The system outputs IPA transcriptions via a 52-symbol phoneme inventory that explicitly represents palatalized variants and two allophonic realisations of the Ukrainian /в/. Word stress position is not a rule dependency but can be supplied externally to disambiguate context-sensitive rules (e.g., where the stressed vowel changes the left context for a consonant mutation).
The lexical stress model is a fine-tuned ByT5 encoder-decoder (Zhu et al., 2022) trained on approximately 135,000 sentences from the Voice of America Ukrainian corpus. Because no manually annotated training data existed, pseudo-labels were generated by running a Wav2Vec2 ASR model configured to output stress marks, followed by cleanup with the Ukrainian Word Stress tool and the Ukrainian Accentor fallback. Sentences are chunked to at most 150 characters before inference to mitigate length degradation in the encoder-decoder architecture. The model was trained for 10 epochs at learning rate 0.0002, achieving 0.58% CER on the training split.
A hybrid inference strategy combines the ByT5 model with Ukrainian Word Stress (OnAmbiguity.Skip): dictionary predictions are used when disambiguation is possible; the neural model fills in the rest.
The benchmark dataset consists of 1,026 sentences manually annotated by a native speaker. To ensure heteronym coverage, 288 additional sentences were generated to provide alternate stress variants for contextually ambiguous words. The dataset contains 640 unique words with stress ambiguity and 1,005 out-of-vocabulary words, deliberately retained to reflect real-world language conditions.
Key Results
On phonemization, the rule-based system achieves 1.23% WER on a manually constructed dataset covering diverse phonological patterns, and 3.07% WER on a VESUM dictionary-derived dataset (excluding controversial transcriptions). A naive letter-to-phoneme baseline scores 48.75% WER on the same data, confirming that explicit rule design is essential for Ukrainian (§4.4, Table 4).
On lexical stress prediction, the standalone ByT5 G2P model reaches 87.7% word-level accuracy and 35.3% sentence-level accuracy, outperforming the word-level Ukrainian Accentor neural baseline (73.2% / 16.6%) and the Ukrainian Accentor Transformer (83.4% / 26.9%). The hybrid combination (ByT5 + Ukrainian Word Stress Skip) yields the best overall performance: 92.5% word-level accuracy and 52.0% sentence-level accuracy (§3.2.4, Table 2).
Contextual disambiguation of heteronyms remains challenging: the hybrid model achieves 61.0% ambiguous word accuracy, compared to 64.3% for the Ukrainian Word Stress (First) dictionary strategy, which benefits from POS-based disambiguation and always falls back to a dictionary entry rather than leaving words unlabelled.
Novelty Assessment
The primary contributions are an engineering integration (combining a new rule-based phonemizer with a fine-tuned ByT5 stress model) and a dataset contribution (the first publicly available sentence-level benchmark for Ukrainian lexical stress). Neither component is architecturally novel: ByT5 is an existing byte-level encoder-decoder, and rule-based G2P systems are well-established. The novelty is in the careful rule engineering for Ukrainian phonology and the data pipeline for generating synthetic stress annotations from broadcast speech. The benchmark dataset is the most reusable output, providing a standardised basis for future Ukrainian TTS preprocessing research.
Field Significance
Moderate — This paper addresses a genuine gap in Ukrainian TTS infrastructure: the absence of open, linguistically accurate preprocessing tools and evaluation resources. The rule-based phonemizer and benchmark dataset provide reusable components for a low-resource Slavic language with non-trivial phonology. The contribution is engineering and data-curation rather than methodological, and its impact is scoped to Ukrainian TTS specifically.
Claims
-
supports: Context-aware sentence-level neural models outperform word-level neural baselines for lexical stress prediction in morphologically complex languages.
Evidence: ByT5 fine-tuned on sentence-level synthetic data achieves 87.7% word-level accuracy, compared to 73.2% for the word-level Ukrainian Accentor, by incorporating broader sentence context during inference. (§3.2.4, Table 2)
-
complicates: Neural models do not surpass dictionary-based systems on contextual stress disambiguation even when trained at the sentence level.
Evidence: ByT5 G2P achieves 58.1% ambiguous word accuracy vs. 64.3% for Ukrainian Word Stress (First); the dictionary system’s POS-based disambiguation and guaranteed fallback remain advantageous for heteronym resolution. (§3.2.4, Table 2)
-
supports: Hybrid pipelines combining neural and dictionary-based stress prediction outperform either method alone for low-resource languages.
Evidence: Combining ByT5 G2P with Ukrainian Word Stress (OnAmbiguity.Skip) achieves 92.5% word-level and 52.0% sentence-level accuracy, exceeding both standalone neural (87.7%, 35.3%) and dictionary-only (88.7%, 41.5%) results. (§3.2.4, Table 2)
-
supports: Carefully designed rule-based G2P systems can achieve very low word error rates for languages with relatively transparent orthographies, substantially outperforming naive mapping.
Evidence: The rule-based Ukrainian phonemizer reaches 1.23% WER on a constructed phonologically diverse dataset, vs. 48.75% WER for a naïve letter-to-phoneme baseline, with errors concentrated in abbreviations and complex consonant clusters. (§4.4, Table 4)
Limitations and Open Questions
Warning
The phonemizer does not handle abbreviations or numerical expressions, and the stress prediction model is trained entirely on automatically generated pseudo-labels with no manually verified training sentences. Both limitations constrain applicability to controlled text in standard Ukrainian orthography.
The stress model’s contextual disambiguation is limited by sparse heteronym representation in the synthetic training corpus, which relies on an ASR-based annotation pipeline with inherent labelling errors. The evaluation benchmark contains only 1,026 sentences, making sentence-level accuracy estimates noisy. Neither system accounts for regional dialects or non-standard language varieties. The phonemizer operates at word level and cannot handle sentence-level phenomena such as cross-word assimilation. Full-text phonemization evaluation is deferred to future work.
Wiki Connections
- Prosody Control — this paper directly addresses lexical stress prediction, an explicit mechanism for controlling stress placement independently of the TTS acoustic model.
- Multilingual TTS — the phonemizer and stress model extend TTS preprocessing coverage to Ukrainian, a low-resource Slavic language with non-deterministic stress.
- Evaluation Metrics — the paper introduces the first publicly available sentence-level benchmark dataset for Ukrainian lexical stress prediction, providing a standardised evaluation framework for future systems.