Source-linked AI summary
Towards Stress-Aware Sentence-Level Filipino G2P With Weakly-Supervised ByT5 Fine-Tuning
Lorenz Bernard Marqueses, Paulo Grane Gabriel Silva, Chastine Cabatay, Ericson Adler Tan, Ann Franchesca Laguna
TL;DR
Filipino G2P becomes more challenging when lexical stress requires sentence-level context, yet sentence-level phoneme data are limited. The paper develops ByT5-based sentence-level models using phoneme-labeled datasets and finds them effective for Filipino G2P, with potential for homograph disambiguation.
Problem
Lexical stress complicates Filipino G2P, while stress-aware sentence-level modeling remains mostly unexplored and sentence-level phoneme data are limited.
Method
The paper constructs phoneme-labeled Filipino datasets, including LLM-assisted synthetic data guided by Wiktionary, and fine-tunes ByT5 for sentence-level G2P.
Results
ByT5 is effective for sentence-level Filipino G2P, and additional training data generally improve performance across PER, CER, and PFER.
Takeaways & Limitations
The approach has potential for Filipino homograph disambiguation and motivates larger, cleaner datasets plus stress-focused modeling in future work.
Takeaways & Limitations
More rigorous testing of the hypotheses about synthetic-data effects is left to future work.
Abstract
from arXiv · showhide
Grapheme-to-phoneme conversion (G2P) refers to the task of converting a sequence of graphemes to a corresponding sequence of phonemes. While Filipino G2P is fairly straightforward due to its shallow orthography, the inclusion of prosodic features such as stress adds a layer of complexity that requires sentence-level context instead of single-word inputs. However, sentence-level data for Filipino typically do not include phoneme transcriptions, posing a challenge for training G2P models. As such, we investigate how to obtain sentence-level phoneme data for Filipino using available data and compare the resulting models with multilingual word-level G2P as well as measure how accurately they predict stress marker position for Filipino. We propose fine-tuning a ByT5-based model, pre-trained on multilingual word-level G2P data, on three sentence-level G2P datasets annotated with an LLM-assisted pipeline guided by data from Wiktionary. This approach produces models that perform well on the G2P task, achieving at best around 0.54% PER and 2.50% CER, a significant decrease compared to base model PER at around 19.74%, on a manually-corrected test set. The model is able to correctly classify most of the main stress classes in Filipino, but struggles particularly with malumi words. We show that a ByT5-based model performs well at sentence-level Filipino G2P and offers strong potential for Filipino homograph disambiguation.
1 Introduction
Filipino G2P is generally facilitated by shallow orthography, but lexical stress creates a sentence-level disambiguation challenge because stress can change meaning without appearing in ordinary spelling.
- G2P converts orthographic symbol sequences into corresponding phonetic symbol sequences.It bridges written and spoken language, including applications such as text-to-speech.
- Filipino’s shallow orthography makes pronunciation generally map directly to spelling.For example, Filipino pronunciation differs from the less direct English pronunciation of mare.
- Lexical stress can change a Filipino word’s meaning even when its spelling remains unchanged.The word bukas can mean “tomorrow” or “open” depending on stress placement.
- Because stress markers are usually omitted from orthography, correct pronunciation requires surrounding phrase or sentence context.For example, context identifies the pronunciation of bukas in bukas ang pinto as “open.”
2 Related work
G2P has progressed from manually crafted rules and statistical sequence models toward transformer architectures, including byte-level ByT5 models that support multilingual and prosodic tasks.
- 2.1 Approaches to G2P: Traditional G2P systems relied on manually crafted phonetic rules, while statistical systems learned sequence patterns with less dependence on explicit rules.Examples include Epitran, Sequitur G2P, and Phonetisaurus.
- 2.1 Approaches to G2P: Modern G2P generally favors transformer architectures because they generalize better across languages and outperform earlier attentionless recurrent approaches.T5 frames G2P as text generation and can use multilingual pretraining when a language-specific model is unavailable.
- 2.2 Byte-level modeling: ByT5 processes raw UTF-8 bytes directly, avoiding complex subword tokenizers and reallocating parameters to dense transformer layers.This supports a learned soft vocabulary for character-to-phoneme mapping.
- 2.2 Byte-level modeling: ByT5 has also been used for accent placement, making its prosodic modeling relevant to Filipino stress restoration.Its direct handling of Unicode combining marks is described as transferable to Filipino’s stress-accent system.
- 2.2 Byte-level modeling: Stress-aware sentence-level Filipino G2P remains largely unexplored despite these byte-level modeling precedents.
3 Methodology
The methodology builds Filipino phoneme resources from Wiktionary-derived word pronunciations and LLM-assisted sentence annotations, then fine-tunes a byte-level encoder-decoder model and evaluates it on manually corrected data.
- 3 Methodology: The task maps a sentence’s word sequence w to a phoneme sequence p using a learned function f: w → p.Training therefore requires both word-level and sentence-level phoneme data capturing Filipino phonetics.
- 3.1 Word-level phoneme data: The word-level resource begins with WikiPron’s Filipino pronunciation dictionary and preserves stress markers from Wiktionary.The initial scrape contains 29,532 unique word-pronunciation pairs.
- 3.1 Word-level phoneme data: Words are classified as semantically ambiguous only when pronunciations span at least two stress classes and correspond to distinct Wiktionary senses.After collapsing duplicates, 1,511 words meet these criteria; remaining entries are treated as non-ambiguous.
- 3.1 Word-level phoneme data: The final phoneme inventory replaces rare or nonstandard scraped symbols with symbols from a Filipino inventory.The combined inventory is presented for comparison with prior Filipino phoneme sets.
- 3.2 Sentence-level phoneme data: An LLM-assisted pipeline generates sentence-level phoneme transcriptions using WikiPron root-word information to guide pronunciation of inflected forms.Because the generated data are noisy, the training setup is weakly supervised.
- 3.2 Sentence-level phoneme data: The sentence datasets are produced from Tatoeba and NewsPH-NLI through LLM processing, filtering, and train-validation-test splitting.Tatoeba yields 5,662 training samples, while processed NewsPH-NLI data yield 22,178 filtered samples overall before splitting.
- Synthetically generated datasets: Synthetic datasets include naive generation and Wiktionary-guided generation, with the latter conditioning sentences on possible word definitions and stress classes.The naive dataset contains 27,494 samples, while the Wiktionary-guided dataset contains 59,252 samples.
- Evaluation: A gold-standard evaluation set contains 1,173 manually corrected transcriptions sampled from the Tatoeba and synthetic datasets.NewsPH-NLI is excluded because its automated transcriptions require relatively extensive manual correction.
4 Experimental setup
The experiments evaluate Filipino G2P with PER, CER, and PFER, while also classifying stress patterns from phoneme transcriptions. Training uses fixed optimization settings, checkpoint ensembling, and an explicit exclusion of the secondary mariin stress phenomenon.
- Evaluation metrics: PER and CER measure phoneme- and character-level transcription errors, with CER additionally including whitespace and stress markers.Both metrics use Levenshtein distance normalized by the total number of phonemes or characters.
- Evaluation metrics: PFER measures articulatory-feature similarity, assigning less severe penalties to phoneme substitutions with similar features.Feature edit costs are computed over PanPhon’s fixed 24-dimensional feature space.
- Stress classification: Filipino stress classification identifies four main patterns—malumay, mabilis, maragsa, and malumi—using stress position and glottal-stop presence.Words that are completely unstressed are classified as none.
- Training configuration: All models use Adafactor with a 3 × 10^-4 learning rate, effective batch size 64, and 10 training epochs.The setup uses warmup scheduling, validation-based checkpoint selection, and checkpoint ensembling across the three best checkpoints.
- Stress classification: The experiments exclude mariin, a secondary stress phenomenon associated with imperfect and contemplative verb aspects, to focus on primary stress.This exclusion follows the distinction between the four primary stress classes and the secondary phenomenon.
5 Results and discussion
Fine-tuned models substantially outperform the base and rule-based baselines, while additional data generally improves error rates. Stress classification is strongest overall for the main classes but weaker for malumi and nonstandard words, and word-level embeddings show limited stress separation.
- Training dataset ablation tests: All fine-tuned models greatly outperform CharsiuG2P and Epitran, although their CER comparison is limited because those baselines do not predict stress markers.The ablation evaluates PER, CER, and PFER on the manually corrected test set.
- Training dataset ablation tests: Adding more training data generally lowers PER, CER, and PFER, while the Wiktionary-guided synthetic set reduces CER by about 0.3%.The Wiktionary-guided addition improves CER relative to the preceding dataset combination, but not all phonetic metrics significantly.
- Wilcoxon signed-rank tests: All pairwise dataset comparisons are significant except combinations (5) and (6), whose PER and PFER differences are nonsignificant while CER differs significantly.For (5) versus (6), PER and PFER have p = 1.0000, whereas CER has p = 0.0015.
- Stress classification: The best model correctly classifies 83% to 89% of words across the four main stress classes, but performs worst on malumi and misclassifies about 30% of nonstandard words as malumay.Evaluation excludes non-ambiguous words and therefore focuses on context-dependent pronunciation choices.
- Word-level embeddings: Word-level embeddings show weak clustering and heavy overlap across stress classes, despite strong separation for unstressed words and small streaks for mabilis.Mean-pooled byte embeddings may obscure stress information, or the model may not represent stress location as a separable internal feature.
6 Conclusion and future work
The study finds ByT5 effective for sentence-level Filipino G2P, with potential for homograph disambiguation, while identifying data scale, cleaning, and stress modeling as priorities for future work.
- ByT5 is effective for sentence-level Filipino G2P and has potential for Filipino homograph disambiguation.
- The paper introduces phoneme-labeled datasets from Tatoeba, NewsPH-NLI, and an LLM-assisted synthetic pipeline.
- Future work should use larger, more diverse datasets and clean training and evaluation samples to ensure correct stress placement.
- Future directions include multi-task learning for stress-location prediction and fusion with contextual models such as Tagalog-pretrained RoBERTa.
- A Hybrid TagRoBERTa + Conformer pipeline with an auxiliary homograph loss head may address homograph ambiguity, agglutinative verb shifts, and stress representation.