Paper deep dive
Phonemizing User-Generated Text: A Benchmark, Taxonomy, and Compositional Approach
MinJu Jeon, Younghan Park, Han Sung Park, Jong-Hwan Kim, Dong-Jin Kim, Hoyeon Lee
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/24/2026, 4:50:54 AM
Summary
This paper introduces UGTPhon, the first grapheme-to-phoneme (G2P) benchmark for user-generated text (UGT) in English, Vietnamese, and Korean. It highlights a significant performance gap in existing G2P models and LLMs when processing non-canonical UGT forms (e.g., abbreviations, slang) compared to canonical text. The authors propose a compositional G2P baseline that uses retrieval-augmented canonical-form hints and staged decoding to improve phonemization accuracy, demonstrating that explicit modeling of canonical-form inference significantly reduces errors.
Entities (10)
Relation Signals (8)
UGTPhon → supportslanguage → Vietnamese
confidence 99% · UGTPhon, the first G2P benchmark for UGT in English, Vietnamese, and Korean
UGTPhon → supportslanguage → Korean
confidence 99% · UGTPhon, the first G2P benchmark for UGT in English, Vietnamese, and Korean
UGTPhon → supportslanguage → English
confidence 99% · UGTPhon, the first G2P benchmark for UGT in English, Vietnamese, and Korean
User-Generated Text → requires → Grapheme-to-Phoneme
confidence 95% · Text-to-speech systems increasingly process user-generated text (UGT)... whose pronunciation must be inferred
UGTPhon → measures → PER
confidence 90% · reaching up to 66.8 PER points... We report Phoneme Error Rate (PER, %)
Qwen2.5-0.5B → usedin → UGTPhon
confidence 90% · Across matched ByT5 and Qwen2.5-0.5B backbones... on UGTPhon
ByT5 → usedin → UGTPhon
confidence 90% · Across matched ByT5 and Qwen2.5-0.5B backbones... on UGTPhon
Epitran → evaluatedon → UGTPhon
confidence 85% · Rule-based G2P Epitran... Table 3: Comparison with existing G2P models on UGTPhon.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Text-to-speech systems increasingly process user-generated text (UGT) such as ppl and imo, whose pronunciation must be inferred from the canonical rather than surface form. We introduce UGTPhon, the first grapheme-to-phoneme (G2P) benchmark for UGT in English, Vietnamese, and Korean, together with an inference-grounded taxonomy for fine-grained diagnosis. Existing G2P models and frontier LLMs exhibit a systematic canonical-to-non-canonical performance gap, reaching up to 66.8 PER points. As a benchmark baseline, we propose a simple compositional G2P approach that incorporates canonical-form evidence through exact-match lookup and staged decoding. Across matched ByT5 and Qwen2.5-0.5B backbones, explicit canonical-form modeling consistently reduces non-canonical G2P errors. The 0.5B variant also performs competitively with much larger few-shot frontier LLMs, highlighting the benefit of explicitly modeling canonical-form inference for UGT phonemization.
Tags
Links
- Source: https://arxiv.org/abs/2609.27205v1
- Canonical: https://arxiv.org/abs/2609.27205v1
Trouble viewing inline? Open PDF directly →
Full Text
82,570 characters extracted from source content.
Expand or collapse full text
Phonemizing User-Generated Text: A Benchmark, Taxonomy, and Compositional Approach MinJu Jeon †thanks: Work performed during an internship at NAVER Cloud. Affiliation: NAVER Cloud Affiliation: Hanyang University Younghan Park Affiliation: Carnegie Mellon University Han Sung Park Affiliation: Georgia Institute of Technology Jong-Hwan Kim Affiliation: NAVER Cloud Dong-Jin Kim †thanks: Corresponding authors. Affiliation: Hanyang University Hoyeon Lee22footnotemark: 2 Affiliation: NAVER Cloud Abstract Text-to-speech systems increasingly process user-generated text (UGT) such as ppl and imo, whose pronunciation must be inferred from the canonical rather than surface form. We introduce UGTPhon, the first grapheme-to-phoneme (G2P) benchmark for UGT in English, Vietnamese, and Korean, together with an inference-grounded taxonomy for fine-grained diagnosis. Existing G2P models and frontier LLMs exhibit a systematic canonical-to-non-canonical performance gap, reaching up to 66.866.8 PER points. As a benchmark baseline, we propose a simple compositional G2P approach that incorporates canonical-form evidence through exact-match lookup and staged decoding. Across matched ByT5 and Qwen2.5-0.5B backbones, explicit canonical-form modeling consistently reduces non-canonical G2P errors. The 0.5B variant also performs competitively with much larger few-shot frontier LLMs, highlighting the benefit of explicitly modeling canonical-form inference for UGT phonemization. 1 Introduction Grapheme-to-phoneme (G2P) conversion Bisani and Ney (2008); Novak et al. (2012); Rao et al. (2015) maps a word’s written form to its phonemic transcription. It serves as the essential front-end of text-to-speech (TTS) systems Ren et al. (2019); Ploujnikov and Ravanelli (2022), where synthesis quality depends directly on the pronunciations it produces Taylor (2005); Van Esch and Sproat (2017). Existing G2P approaches, including joint-sequence models Bisani and Ney (2008), transformer architectures Yolchuyeva et al. (2019), and pretrained byte-level models Zhu et al. (2022), achieve robust performance on canonical lexicons such as IPA-Dict Doherty (2016) and WikiPron Lee et al. (2020). TTS pipelines, therefore, treat phonemization as a reliable upstream module Kim et al. (2021); Shen et al. (2018), assuming input that follows standard orthography. Figure 1: Conventional G2P models (red, left) read non-canonical words (ppl, luv, imo) letter by letter, producing letter names. Accurate phonemization (green, right) therefore requires recovering the canonical form (people, love, in my opinion) before generating phonemes. However, this assumption breaks down in practice. As TTS systems increasingly process conversational and web-based text Lee and Ashok (2022), input is dominated by user-generated text (UGT) Plank (2016); Han and Baldwin (2011): a mixture of canonical words (e.g., love, people) and non-canonical forms like luv and ppl. These non-canonical forms follow systematic patterns recurring across languages such as English (EN), Vietnamese (VI), and Korean (KO): abbreviations (EN: lmao, VI: vcl, KO: ㅇㄱㄹㅇ), orthographic lengthening (cooool, đi, ㅋ), phonetic substitutions (gr8, wá, 머해), and slang (lit, gato, 킹받다). Existing G2P models therefore read these forms literally and yield letter-name strings rather than the intended pronunciations (Figure 1 (A)). Prior work has framed UGT as noise (Zhao et al., 2022), but UGT is fundamentally different: imo is a deliberate substitution whose pronunciation must be inferred from its canonical expansion (in my opinion), not its surface form. G2P robustness under UGT remains unexplored. To address this gap, we introduce UGTPhon 11 1 The UGTPhon is publicly available at https://github.com/naver-ai/UGTPHON., the first G2P benchmark for UGT, spanning three typologically diverse languages (English, Vietnamese, and Korean) with the intended canonical form and gold phoneme sequence for every non-canonical word. To support fine-grained diagnosis, we further group non-canonical words by the inference a G2P model must perform: (1) Reconstruct (recover dropped characters, e.g., idk for I don’t know), (2) Repetition (collapse prosodic repetitions, e.g., soooo for so), and (3) Pass-through (the surface already encodes the pronunciation, e.g., luv for love). This inference-grounded design lets a model’s failure pattern indicate which kind of inference is missing, not which inputs it fails on. Non-canonical G2P composes two stages, inferring the canonical form and phonemizing it. Standard end-to-end training collapses these into one, while a naive normalize-then-G2P pipeline over-corrects forms that need no rewriting. We therefore use a simple compositional G2P baseline with three components: Contextualized input encoding marks the target word within its sentence, Retrieval supplies a canonical-form hint from an external datastore, and Staged decoding emits the canonical form as an intermediate target before its phonemes (<Target> GcG_c <Phoneme> PcP_c). On UGTPhon, all baselines suffer degradation on non-canonical words, with the canonical–non-canonical gap reaching up to 66.8 PER points. Further training on UGTPhon substantially lowers non-canonical error rates on English and Vietnamese but leaves a large gap. Our compositional baseline consistently reduces non-canonical G2P error relative to matched ByT5 and Qwen2.5-0.5B backbones across all three languages, while the 0.5B variant is also competitive with much larger few-shot frontier LLMs. We summarize our contributions as follows: • UGTPhon, the first G2P benchmark for user-generated text across English, Vietnamese, and Korean, with an inference-grounded taxonomy for fine-grained diagnosis. • A systematic evaluation showing how canonical and non-canonical G2P errors vary across inference types, model families, and languages. • A simple compositional baseline that injects canonical-form evidence through exact-match lookup and staged decoding, consistently improving matched supervised backbones on UGTPhon. English Vietnamese Korean Type ID Category Gnc→GcG_nc\!→\!G_c tr+dv / te Gnc→GcG_nc\!→\!G_c tr+dv / te Gnc→GcG_nc\!→\!G_c tr+dv / te A Reconstruct A1 Shortening-V pls → please 595/166 đc → được 3,112/327 ㅅㄱ → 수고 451/67 A2 Shortening-O bc → because 313/82 k → không 4,609/496 – – A3 Phrasal (Unp) idk → I don’t know 504/117 mn → mọi người 2,020/225 – – B Repetition B1 Lengthening soooo → so 175/75 đi → đi 475/43 구우웃 → 굿 166/66 B2 Iteration waitwait → wait wait 113/44 xem xem → xem xem 157/38 ㅋ → 크, 키 322/136 C Pass-through C1 Eye Direct luv → love 515/140 zô → vô 2,385/368 어케 → 어케 509/119 C2 Regular droppin → dropping 1,559/405 iu → yêu 2,164/247 걍 → 걍 175/44 C3 Slang/Teen tho → though 903/236 hong → không 2,897/369 막짤 → 막짤 1,205/234 C4 Phrasal (Pro) rofl → rofl 166/62 GATO → GATO 129/39 월클 → 월클 89/56 Table 1: Diagnostic taxonomy of non-canonical UGT in UGTPhon, regrouping prior orthographic taxonomies by the inference G2P must perform: three types ordered by inference demand, with nine fine-grained categories. (A) Reconstruct: recover dropped characters. (B) Repetition: collapse repeats. (C) Pass-through: surface already encodes pronunciation. Each cell: example Gnc→GcG_nc\!→\!G_c and counts (train+dev / test); “–”: absent. English Vietnamese Korean Total Train 3,329 8,597 1,976 13,902 Dev 600 1,129 246 1,975 Test 1,100 1,121 275 2,496 Total 5,029 10,847 2,497 18,373 Table 2: UGTPhon composition by language and split. 2 UGTPhon We propose UGTPhon, the first G2P benchmark for UGT, comprising 18,373 sentences across three typologically diverse languages (Table 2): English (5,029), Vietnamese (10,847), and Korean (2,497). We formalize G2P on UGT (Section 2.1), group non-canonical words by required inference into a diagnostic taxonomy (Section 2.2), and annotate (Gnc,Gc,Pc)(G_nc,G_c,P_c) at the target level within full sentential context (Section 2.3). This supports aggregate evaluation and category-level diagnosis of G2P failures. 2.1 Task Formulation Notation. Let S denote a sentence and wiw_i an annotated target token or span in S, which is either canonical or non-canonical. In Vietnamese, a single annotated target may span multiple whitespace-delimited units. A canonical word has a spelling listed in the standard dictionary (e.g., love, please, không). A non-canonical word GncG_nc is a conventionalized variant of a canonical form arising systematically in UGT (e.g., luv, pls, k). We denote the canonical form of GncG_nc as GcG_c and the corresponding gold IPA phoneme sequence as PcP_c. For each annotated target, we represent the G2P task as a quadruplet (S,Gnc,Gc,Pc)(S,G_nc,G_c,P_c), where the non-canonical form GncG_nc appears as a token or span in sentence S. Canonical form annotation. We adopt the standard formulation of lexical normalization Van Der Goot et al. (2018); Van Der Goot et al. (2021); Nguyen et al. (2024): the mapping Gnc→GcG_nc _c admits 1-to-1 (pls → please), 1-to-N (idk → I don’t know), and N-to-1 (ch ó → chó) mappings. We obtain GcG_c from LexNorm2015 Baldwin et al. (2015), via MultiLexNorm Van Der Goot et al. (2021), for English and from ViLexNorm Nguyen et al. (2024) for Vietnamese. For Korean, we draw GncG_nc from MultiLexNorm++ Buaphet et al. (2026) and KoMultiText Choi et al. (2023), and annotate GcG_c from scratch with native-speaker experts. (procedure in Section A.2). Task definition. Existing G2P benchmarks evaluate word-level conversion in isolation Bisani and Ney (2008); Novak et al. (2012); Rao et al. (2015). This is insufficient for UGT: a non-canonical surface like ig can normalize to I guess or Instagram depending on the sentence. UGTPhon therefore situates every non-canonical word within its original sentence S. Given GncG_nc and S, the task is to predict the gold phoneme sequence PcP_c derived from the intended canonical form GcG_c, which remains latent during evaluation. Evaluation is at the target level for precise error attribution, while conditioning on the full sentence. 2.2 Diagnostic Taxonomy of Non-Canonical Words Non-canonical surfaces in UGT vary in the inference required for G2P: some can be read directly, others require recovering dropped characters, and others require collapsing repeated units. Existing UGT taxonomies Van Der Goot et al. (2018) group surfaces by orthographic transformation, which is suited to lexical normalization but not phonemization. UGTPhon instead regroups these surfaces by the inference required for phonemization (Table 1), so that failure patterns localize which kind of inference a model lacks. Coarse-grained types. We define three coarse types, ordered from the most to the least demanding inference. (A) Reconstruct. Surface content required for pronunciation has been omitted and must be recovered, possibly as a multi-word expansion (e.g., idk → I don’t know). (B) Repetition. Characters or units have been repeated for prosodic emphasis Brody and Diakopoulos (2011) and must be collapsed (e.g., soooo → so). (C) Pass-through. The surface already preserves the intended pronunciation and can be read directly (e.g., luv → love). Fine-grained categories. Within each coarse type, we distinguish 2–4 fine-grained categories reorganized from prior UGT taxonomies Van Der Goot et al. (2018). Reconstruct splits by what is omitted (vowels A1, other characters A2, full phrases A3). Repetition splits by what is repeated (characters B1, units B2). Pass-through splits by how pronunciation is preserved in the surface (phonetic respelling C1, productive variation C2, lexicalized slang C3, pronounceable acronyms C4). Examples and per-language counts are shown in Table 1. Korean Reconstruct collapses to A1, since it consists almost entirely of initial-consonant abbreviations (e.g., ㄱㅅ→ 감사). Figure 2: Three approaches to non-canonical G2P. (A) Vanilla G2P reads words literally. (B) A lexical normalization model rewrites the input but overcorrects canonical words, propagating errors downstream. (C) Our retrieval-augmented compositional G2P baseline injects a Gnc→GcG_nc _c canonical form hint and emits GcG_c as an intermediate before the phonemes. 2.3 Data Construction We construct UGTPhon in three stages: source corpora selection, phoneme annotation, and category labeling. We preserve the raw UGT sentence in the released text field. During phoneme construction, non-lexical elements such as punctuation and numbers are dropped or verbalized as appropriate, while unphonemizable items may remain as placeholders. Full details, licenses, and annotation procedures are provided in Appendix A. Source corpora. For English, GncG_nc–GcG_c pairs are derived from LexNorm Baldwin et al. (2015), via MultiLexNorm Van Der Goot et al. (2021). For Vietnamese, they are derived from ViLexNorm Nguyen et al. (2024). For Korean, GncG_nc is drawn from MultiLexNorm++ Buaphet et al. (2026) and KoMultiText Choi et al. (2023), and GcG_c is annotated from scratch (Section A.2). In addition, the English and Vietnamese subsets include 175 and 385 human-curated augmented UGT sentences, respectively, generated from existing UGT patterns. We preserve the natural resource asymmetry across languages rather than down-sampling, ensuring sufficient samples per category in every split. Since the source corpora include UGT from social media, UGTPhon inherits some toxic content; we report a sentence-level moderation analysis in Appendix B. Phoneme annotation. Target phonemes PcP_c are derived from GcG_c via IPA-Dict Doherty (2016), an expert-curated grapheme-to-IPA lexicon, and verified by native-speaker experts for language-specific phenomena (English heteronyms, Vietnamese tones, Korean phonological rules). For canonical forms GcG_c not covered by the lexicon (e.g., recent loanwords or proper nouns), experts provide PcP_c from scratch and cross-verify. Inter-annotator agreement reaches Cohen’s κ=0.86κ=0.86 (English) and 0.870.87 (Vietnamese), both in the almost-perfect range Landis and Koch (1977), with Korean at 0.750.75 (substantial) reflecting variability in IPA transcription rather than disagreement on the underlying pronunciation. Per-category agreement and OOV statistics are reported in Section A.4. Category labeling. Each non-canonical word is assigned to one of the nine categories (Table 1) through manual annotation by native speakers, with second-pass verification and disagreements reconciled through discussion. Operational definitions and full guidelines are in Section A.3 (Table 7). 3 Method Two conventional approaches to non-canonical G2P lack joint supervision of the normalization-phonemization composition (Figure 2). Vanilla G2P (Figure 2 (A)) phonemizes the surface directly, leaving Gnc→GcG_nc _c unsupervised. Lexical normalization + G2P (Figure 2 (B)) decouples the stages but over-corrects canonical-recoverable surfaces. We use a retrieval-augmented compositional baseline (Figure 2 (C)) that jointly supervises both stages within a single architecture. Our approach comprises three components. First, Contextualized input encoding (Section 3.1) marks the target within its sentential context. Then, Retrieval-augmented canonical-form prompting (Section 3.2) supplies a Gnc→GcG_nc _c hint through an exact-match lookup when the surface form occurs in a training-derived datastore. Finally, Staged decoding (Section 3.3) emits GcG_c before PcP_c, binding the retrieved hint to prediction. 3.1 Contextualized Input Encoding For each annotated target token or span wiw_i in sentence S, we construct the input string by prepending a language tag and surrounding wiw_i with pipe markers: <lang>w1w2⋯|wi|⋯wn. <lang>\ w_1\,w_2\,·s\, |\,w_i\, |\,·s\,w_n. This format treats wiw_i as a single annotated target span, even when its surface or canonical form spans multiple whitespace-delimited units (e.g., idk → I don’t know). Each target is processed as an independent forward pass, so multiple non-canonical words do not accumulate cross-word errors during decoding Yoon et al. (2023). This input format applies uniformly across training and inference, and across different backbone families: ByT5-small Xue et al. (2022) and Qwen 2.5-0.5B Qwen et al. (2025). Since the | symbol is a regular character in both vocabularies, no special tokens are added. Our approach is therefore backbone-agnostic and does not require backbone-specific tokenization. We use ByT5-small as the default backbone for its byte-level input, which handles UGTPhon’s diverse scripts without language-specific tokenizers. 3.2 Retrieval-Augmented Canonical-Form Inference Training a model to map GncG_nc directly to PcP_c provides no explicit supervision for the Gnc→GcG_nc _c normalization stage: the surface form bears little phonemic correspondence to the target (idk vs. /aI doUnt noU/). We augment the input with an optional canonical-form hint retrieved through a simple exact-match key–value lookup over a training-derived datastore. Retrieval datastore. We construct the datastore D as a key–value map from training-set (Gnc,Gc)(G_nc,G_c) pairs, with GncG_nc as the key and GcG_c as the value. For each target wiw_i in S, we retrieve its canonical hint from D: prefix(wi)=<Hint>wi→[wi]if wi∈,<Miss>otherwise.prefix(w_i)= cases <Hint>\ w_i [w_i]&if w_i ,\\ <Miss>&otherwise. cases (1) On a miss, the <Miss> marker explicitly marks retrieval failure as a learnable condition, prompting the model to rely on sentence context for normalization. Optional LLM-based datastore expansion. New non-canonical forms appear continuously, and the training split cannot exhaust the space of variants seen at test time. To extend coverage, we prompt GPT-5.4-mini Singh et al. (2025) with seed pairs from D to generate additional (Gnc,Gc)(G_nc,G_c) candidates following the same UGT category patterns. Seeds and all filtering are restricted to the training split. We discard candidates whose GncG_nc matches a standard word in the train or dev split, to avoid spuriously rewriting canonical inputs, or where Gc=GncG_c=G_nc, which carries no normalization signal. Prompt details and coverage statistics are provided in Appendix C. Hit and Miss prefixes. The two special tokens introduced in Equation 1, <Hint> and <Miss>, are added to the vocabulary. Prepending the retrieval prefix to the contextualized input from Section 3.1 produces the following inputs: Hit: <lang> <Hint> wi→Gc,iw_i→G_c,i w1w2⋯w_1\,w_2\,·s\,|wiw_i|⋯wn\,·s\,w_n Miss: <lang> <Miss> w1w2⋯w_1\,w_2\,·s\,|wiw_i|⋯wn\,·s\,w_n The <Miss> marker also distinguishes a non-canonical word absent from D from a canonical input that requires no retrieval. English Vietnamese Korean Model PERipaPER_ipa ↓ PERcPER_c ↓ PERnc⟨Δ⟩PER_nc\, ↓ PERipaPER_ipa ↓ PERcPER_c ↓ PERnc⟨Δ⟩PER_nc\, ↓ PERipaPER_ipa ↓ PERcPER_c ↓ PERnc⟨Δ⟩PER_nc\, ↓ Rule-based G2P Epitran (–) 33.4 13.4 45.1 ⟨ +31.7⟩ 32.7 34.7 75.1 ⟨ +40.4⟩ 47.0 37.5 50.7 ⟨ +13.2⟩ viG2P (–) – – – 25.4 27.3 78.4 ⟨ +51.1⟩ – – – Supervised G2P (trained on IPA-Dict canonical lexicon) Transformer (2019) (2.0M) 17.4 12.9 44.8 ⟨ +31.9⟩ 15.3 5.8 68.4 ⟨ +62.6⟩ 7.5 9.5 54.8 ⟨ +45.3⟩ ByT5-tiny (2022) (7.3M) 17.0 18.8 46.4 ⟨ +27.6⟩ 15.2 6.0 71.9 ⟨ +65.9⟩ 7.2 6.8 51.1 ⟨ +44.3⟩ GBERT (2022) (8.0M) 21.5 17.6 45.0 ⟨ +27.4⟩ 24.0 30.8 93.0 ⟨ +62.2⟩ 11.7 16.8 45.9 ⟨ +29.1⟩ ByT5 (2022) (300M) 13.5 10.5 46.6 ⟨ +36.1⟩ 15.2 5.9 72.7 ⟨ +66.8⟩ 6.7 4.2 46.2 ⟨ +42.0⟩ Supervised G2P (further trained on UGTPhon) Transformer (2.0M) 18.8 10.3 32.7 ⟨ +22.4⟩ 17.8 5.9 33.7 ⟨ +27.8⟩ 7.5 5.8 38.7 ⟨ +32.9⟩ ByT5-tiny (7.3M) 30.1 9.7 30.8 ⟨ +21.1⟩ 17.0 7.7 43.6 ⟨ +35.9⟩ 8.1 4.0 39.0 ⟨ +35.0⟩ GBERT (8.0M) 25.6 7.8 50.5 ⟨ +42.7⟩ 50.3 5.6 60.4 ⟨ +54.8⟩ 8.5 5.8 31.6 ⟨ +25.8⟩ ByT5 (300M) 15.9 5.2 29.8 ⟨ +24.6⟩ 17.3 7.7 43.2 ⟨ +35.5⟩ 7.2 4.3 47.5 ⟨ +43.2⟩ Ours (ByT5) (300M) 17.8 4.7 22.5 ⟨ +17.8⟩ 15.9 5.9 26.7 ⟨ +20.8⟩ 9.0 5.1 32.0 ⟨ +26.9⟩ Ours (Qwen 2.5-0.5B) (494M) 19.5 10.2 14.3 ⟨ +4.1⟩ 15.5 4.0 20.0 ⟨ +16.0⟩ 8.2 6.0 32.4 ⟨ +26.4⟩ Table 3: Comparison with existing G2P models on UGTPhon. PER (%) on canonical (PERcPER_c) and non-canonical (PERncPER_nc) words, with the within-input gap Δ=PERnc−PERc =PER_nc-PER_c. Color encodes Δ : ≥30≥ 30 p (large gap), 1010–3030 p (medium), <10<10 p (small). Bold: best per column. 3.3 Staged Decoding To bind the retrieved hint to prediction, we make the canonical form an explicit intermediate target during decoding. The model produces the normalization output GcG_c before the phoneme sequence PcP_c in a single autoregressive trajectory, with the target sequence structured as follows: Non-canonical: <Target>Gc⏟normalization<Phoneme>Pc⏟phonemization, <Target>\,G_c_normalization\, <Phoneme>\,P_c_phonemization, Canonical: <Phoneme>Pc. <Phoneme>\,P_c. For canonical words, the model bypasses the normalization stage and emits PcP_c directly, aligning with conventional G2P training. For non-canonical words, it generates GcG_c as a learned intermediate before PcP_c, providing explicit supervision for both the normalization stage (via GcG_c) and the phonemization stage (via PcP_c). The model is trained end-to-end with cross-entropy loss over the full target sequence, without auxiliary objectives or additional loss terms. Together, the three components make canonical-form modeling explicit within a single G2P architecture, supervising normalization and phonemization without sacrificing canonical-token accuracy. 4 Experiments Using the taxonomy of Section 2.2, we quantify the canonical-to-non-canonical gap on UGTPhon, decompose it by category, and compare our model against two alternatives: a normalize-then-phonemize pipeline and frontier LLMs. 4.1 Experimental Setup Evaluation protocol. Models take the full sentence as input and predict the marked target word. We report Phoneme Error Rate (PER, %) on UGTPhon’s test split as PERcPER_c (canonical) and PERncPER_nc (non-canonical), with Δ=PERnc−PERc =PER_nc-PER_c. For lexicon-only supervised baselines, we additionally report PERipaPER_ipa on the IPA-dict Doherty (2016) lexicon to track canonical coverage. Baselines. We compare our model against task-relevant G2P, normalization, and frontier-LLM baselines. Rule-based G2P: Epitran Mortensen et al. (2018) and viG2P Tri et al. (2020). Supervised G2P: Transformer Yolchuyeva et al. (2019), ByT5/ByT5-tiny Zhu et al. (2022), and GBERT Dong et al. (2022), each evaluated in two training regimes: lexicon-only (trained on canonical IPA-Dict) and fine-tuned (further trained on UGTPhon). Frontier LLMs: GPT-4o, GPT-5.4-mini (Singh et al., 2025), Gemini-2.5-flash, Gemini-3-flash, and Claude-Sonnet-4.6, evaluated under the nominal K=9K=9 in-context setting described in Appendix E, following prior G2P-with-LLM work (Han et al., 2024). Task-relevant baselines: ICKR Han et al. (2024), R-G2P Zhao et al. (2022), and lexical-normalization ++ G2P pipelines using MoNoise (Van Der Goot, 2019) or PolyNorm (Wong et al., 2025). The full cross-lingual comparison is reported in Table 10. Implementation. We use ByT5-small as the default backbone, with Qwen2.5-0.5B for backbone ablation. Training uses AdamW (lr=10−4lr=10^-4, batch size=32, 10 epochs) on 4 NVIDIA V100 GPUs. The retrieval datastore D uses the exact-match lookup described in Section 3.2. For the frontier LLMs, we query official APIs under the nominal K=9K=9 in-context setting described in Appendix E. Model identifiers and access dates are listed in Appendix D; the prompt template is provided in Appendix E. Input Retrieval Augmentation Staged Decoding English Vietnamese Korean PERallPER_all ↓ PERcPER_c ↓ PERncPER_nc ↓ PERallPER_all ↓ PERcPER_c ↓ PERncPER_nc ↓ PERallPER_all ↓ PERcPER_c ↓ PERncPER_nc ↓ Word→ ✗ ✗ 7.9 5.2 29.8 14.1 7.7 43.2 17.8 4.3 47.5 Sent→ ✗ ✗ 19.7 (++11.8) 18.0 (++12.8) 30.5 (++0.7) 29.7 (++15.6) 19.2 (++11.5) 70.7 (++27.5) 33.5 (++15.7) 41.6 (++37.3) 31.3 (−-16.2) Sent→ ✗ ✗ 7.6 (−-0.3) 5.4 (++0.2) 25.5 (−-4.3) 15.0 (++0.9) 9.1 (++1.4) 41.9 (−-1.3) 12.0 (−-5.8) 4.7 (++0.4) 28.1 (−-19.4) Sent→ ✗ ✓ 8.9 (++1.0) 4.9 (−-0.3) 30.1 (++0.3) 10.1 (−-4.0) 5.9 (−-1.8) 31.8 (−-11.4) 13.1 (−-4.7) 4.0 (−-0.3) 33.1 (−-14.4) Sent→ ✓ ✗ 7.0 (−-0.9) 5.0 (−-0.2) 22.5 (−-7.3) 9.6 (−-4.5) 5.8 (−-1.9) 29.1 (−-14.1) 12.4 (−-5.4) 5.5 (++1.2) 27.7 (−-19.8) Sent→ ✓ ✓ 6.8 (−-1.1) 4.7 (−-0.5) 22.5 (−-7.3) 9.3 (−-4.8) 5.9 (−-1.8) 26.7 (−-16.5) 13.5 (−-4.3) 5.1 (++0.8) 32.0 (−-15.5) Table 4: Component ablation of our retrieval-augmented G2P model on ByT5-small. Green/Red: gain/regression vs. the Word→ baseline. Retrieval Aug: Gnc→GcG_nc→G_c hints from the datastore. Staged Decoding: <Target> → <Phoneme> emission. Bold: column-best per language (PER, %). 4.2 Main Results All baselines exhibit large canonical-to-non-canonical gaps. Table 3 reports PER for rule-based and supervised baselines on UGTPhon. The within-input gap Δ ranges from +27.4+27.4 to +36.1+36.1 on English, +40.4+40.4 to +66.8+66.8 on Vietnamese, and +13.2+13.2 to +45.3+45.3 on Korean. The pattern persists across rule-based Mortensen et al. (2018), transformer Yolchuyeva et al. (2019), and byte-level Zhu et al. (2022) architectures, and is most severe on Vietnamese, where supervised baselines average Δ≈+64 ≈+64 despite PERcPER_c below 8%8\%. Further training on UGTPhon reduces PERncPER_nc but does not close Δ . Training the supervised baselines on UGTPhon’s training split substantially lowers PERncPER_nc on English (46.6→29.846.6→ 29.8) and Vietnamese (72.7→43.272.7→ 43.2), but the strongest fine-tuned model (ByT5) still shows Δ=+24.6 =+24.6 on English and +35.5+35.5 on Vietnamese. Korean even slightly worsens (46.2→47.546.2→ 47.5), reflecting its Pass-through–dominated composition in Table 1, where the surface already encodes pronunciation and further fine-tuning provides little signal. The remaining Δ shows that further training alone does not eliminate the non-canonical gap, motivating explicit surface-to-canonical modeling. Explicit canonical-form modeling improves matched backbones. Relative to matched ByT5 fine-tuning, our formulation lowers PERncPER_nc from 29.829.8 to 22.522.5 on English, 43.243.2 to 26.726.7 on Vietnamese, and 47.547.5 to 32.032.0 on Korean. The same pattern holds for Qwen2.5-0.5B: 14.8→14.314.8→ 14.3, 23.1→20.023.1→ 20.0, and 77.3→32.477.3→ 32.4, respectively. Three-seed runs show stable English/Vietnamese results, while Korean has higher variance; paired bootstrap tests on the reported checkpoints give confidence intervals excluding zero for all six matched-backbone PERncPER_nc comparisons (Tables 11 and 12). The largest matched-backbone reductions occur on Vietnamese and Korean, while English Qwen improves from 14.814.8 to 14.314.3 PERncPER_nc. 4.3 Analysis Sentence context helps as evidence, not as a generation target. Table 4 shows that generating the full sentence (Sent→ ) hurts PERcPER_c across all three languages, reflecting exposure bias from long autoregressive outputs Yoon et al. (2023) (Korean PERcPER_c rises by 37.337.3 points). Sent→ avoids this trade-off and lowers Korean PERncPER_nc (47.5→28.147.5→ 28.1), since Korean is dominated by pass-through forms that sentence context alone can disambiguate. Vietnamese, by contrast, improves little under Sent→ , motivating the explicit canonical-form evidence that the lookup provides. Exact-match lookup supplies canonical-form evidence; staged decoding encourages the model to use it. On top of Sent→ , retrieval alone reduces PERncPER_nc from 41.941.9 to 29.129.1 on Vietnamese and 25.525.5 to 22.522.5 on English, supplying the surface-to-canonical evidence the surface lacks. Staged decoding alone improves Vietnamese (41.9→31.841.9→ 31.8) but worsens English (25.5→30.125.5→ 30.1) and Korean (28.1→33.128.1→ 33.1): without a retrieval hint, forcing a canonical intermediate invites hallucinated reconstructions. The two together attain the lowest PERncPER_nc on English (22.522.5) and Vietnamese (26.726.7) while keeping PERcPER_c at or below the no-retrieval level. Korean shows no further gain, since its bottleneck has shifted from normalization to phonemization. Figure 3: PERncPER_nc (%) by fine-grained category. Gains concentrate on Reconstruct (A), with smaller margins on Repetition (B) and Pass-through (C). English Model PERcPER_c PERncPER_nc word-ByT5 (no norm.) 5.2 29.8 word-ByT5 + MoNoise (2019) 14.2 18.0 word-ByT5 + GPT-5.4-mini 13.6 18.6 Ours (ByT5) 4.7 22.5 Ours (Qwen 2.5-0.5B) 10.2 14.3 Table 5: Comparison with two-stage LexNorm ++ G2P pipelines on English (PER, %). Improvements track the inference demand of each category. Figure 3 decomposes PERncPER_nc along the taxonomy of Table 1, which predicts where canonical-form modeling should help and where it should add little. The results match. Reconstruct (A) is where the surface and canonical form share the least phonemic evidence (e.g., idk → I don’t know), so our model’s gain is largest: with the Qwen 2.5 (0.5B) backbone, A-average PERncPER_nc drops by 33.433.4 points on English (54.4→21.054.4→ 21.0) and 15.215.2 on Vietnamese (38.3→23.138.3→ 23.1). On English A3 specifically, our model reaches 11.511.5, compared with 66.366.3 for the word-level baseline and 36.836.8 for GPT-4o, yielding the largest separation on a category that requires phrasal canonical-form reconstruction. Repetition (B) is intermediate: the surface encodes the pronunciation but with redundant content that must be collapsed. Our model attains the best PERncPER_nc on English B2 and sharply reduces Vietnamese B2 against frontier LLMs, though the word-level ByT5 baseline remains marginally stronger on Vietnamese B2 (22.522.5 vs. 26.326.3), since the redundant surface still preserves the underlying phonemes. Pass-through (C) is mixed: the surface already encodes the pronunciation, so the canonical-form intermediate is no longer strictly necessary. Where the baseline is near-zero (Korean C4, 1.91.9), our model offers no benefit; where the surface remains non-trivial to phonemize, the gain returns. Our model attains the best PERncPER_nc on nine of twelve Pass-through subcategories, so the cost-benefit depends on residual phonemization difficulty rather than the coarse category label. Figure 4: Shot scaling for frontier LLMs on UGTPhon. PERncPER_nc (%) under the nominal K∈0,9,27K∈\0,9,27\ shot settings; Korean uses 0, 7, and 21 actual in-context examples because A2 and A3 are absent. Our approach outperforms two-stage pipelines without trading canonical for non-canonical accuracy. Table 5 compares our model against two-stage English pipelines as shown in Figure 2 (B) that pair a lexical normalizer with word-level ByT5: MoNoise Van Der Goot (2019) and GPT-5.4-mini Singh et al. (2025). English is the only language where this comparison is possible, since dedicated lexical normalization models exist for English but not for the lower-resource languages we evaluate. The two pipelines reduce PERncPER_nc from 29.829.8 to roughly 1818, but PERcPER_c degrades from 5.25.2 to over 1313. Our model avoids this trade-off: with the ByT5 backbone, PERcPER_c stays at 4.74.7 and PERncPER_nc drops to 22.522.5; with Qwen 2.5-0.5B, PERncPER_nc reaches the lowest value across all systems (14.314.3) while PERcPER_c (10.210.2) remains well below the pipelines. Our approach remains competitive with frontier LLMs. Figure 4 compares our approach against five frontier LLMs under the nominal K∈0,9,27K∈\0,9,27\ shot settings. Across all three languages, our 0.5B model achieves performance competitive with or better than the evaluated frontier LLMs. On English, the strongest frontier model (Gemini-3-flash) reaches PERnc=17.4PER_nc=17.4 under K=27K=27, while our Qwen2.5-0.5B model achieves 14.314.3. On Vietnamese, frontier LLMs exhibit substantially larger errors on several UGT categories, including GPT-5.4-mini on Reconstruct A3 (PERnc=109.6PER_nc=109.6; Figure 3), whereas our model achieves the lowest overall PERncPER_nc of 20.020.0. On Korean, our ByT5 and Qwen2.5-0.5B variants achieve PERncPER_nc of 32.032.0 and 32.432.4, respectively, compared with 42.642.6 for the strongest evaluated frontier baseline (Claude-Sonnet-4.6 under K=27K=27). These results further highlight the difficulty of canonical-form inference for UGT, particularly in Vietnamese and Korean. The training-set datastore alone closes most of the gap to oracle retrieval. Table 6 compares four datastores: no retrieval, training-set entries only, training-set with LLM-generated augmentation, and an oracle that uses gold canonical forms at test time. The training-set datastore alone reaches PERnc=22.7PER_nc=22.7 on English, within 2.12.1 points of the oracle (20.620.6), and 26.726.7 on Vietnamese, within 0.80.8 points of the oracle (25.925.9). Adding LLM-generated (Gnc,Gc)(G_nc,G_c) pairs reduces English PERncPER_nc by only 0.20.2 point and leaves Vietnamese unchanged. The deployable datastore approaches oracle performance without test-time canonical-form access. Datastore D PERncPER_nc (%) EN VI KO No retrieval (∅ ) 30.1 31.8 33.1 Train-only 22.7 26.7 32.4 + LLM Aug 22.5 26.7 32.0 Oracle (not deployable) 20.6 25.9 18.6 Table 6: Retrieval datastore D sensitivity (ByT5). Oracle uses gold canonical forms (not deployable). Downstream TTS-ASR validation. On downstream TTS-ASR evaluations with eSpeak NG and Whisper large-v3 (Radford et al., 2023) (Appendix F), our ByT5 variant attains the lowest WER on English (0.3590.359), while our Qwen2.5-0.5B variant attains the lowest observed WER among predicted-input systems on Vietnamese (0.6400.640). These results are consistent with the token-level PER trends. 5 Conclusion We diagnose user-generated text as a setting where G2P must compose surface-to-canonical inference with phonemization, and introduce UGTPhon, the first G2P benchmark for UGT with an inference-grounded taxonomy across English, Vietnamese, and Korean. Our retrieval-augmented compositional baseline makes canonical-form inference explicit through exact-match hints and staged decoding, reducing non-canonical PER relative to matched ByT5 and Qwen2.5-0.5B backbones across all three languages. The gains are largest on Reconstruct cases, while the 0.5B variant remains competitive with the evaluated frontier LLMs. Together, UGTPhon provides a multilingual benchmark and diagnostic framework for studying canonical-form inference in UGT phonemization. Limitations We include English and Vietnamese TTS-ASR evaluations (Appendix F) showing that lower PER is accompanied by lower word error rate for several systems, but we do not report subjective MOS ratings or evaluate across multiple TTS systems; characterizing the full relationship between PER and TTS quality on UGT is left to future work. UGTPhon spans three typologically diverse languages (alphabetic English, tonal-diacritic Vietnamese, syllabic-block Korean) but does not cover logographic, abjad, or abugida writing systems, where UGT phenomena and the relevant normalization signals may differ. Whether the compositional decomposition transfers to these scripts is an open question. Our retrieval datastore is built from training-set pairs and an LLM-based expansion. UGT vocabulary drifts continuously, so maintaining coverage on emerging non-canonical forms will require periodic datastore refresh; we leave the dynamics of such refresh, and zero-shot extensions of the retrieval mechanism, to future work. Beyond these scope limitations, more accurate UGT phonemization is dual-use: it improves TTS intelligibility on harmful UGT as well as benign UGT, and UGTPhon inherits toxic content from its source corpora, whose moderation profile we report in Appendix B. Acknowledgments This work was partly supported by NAVER Cloud Corporation, the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No. RS-2020-I201373, Artificial Intelligence Graduate School Program, Hanyang University), and the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No. RS-2025-02215122, Development and Demonstration of Lightweight AI Model for Smart Homes). References Baldwin et al. (2015) T. Baldwin, M. De Marneffe, B. Han, Y. Kim, A. Ritter, and W. Xu Shared tasks of the 2015 workshop on noisy user-generated text: twitter lexical normalization and named entity recognition. In Proceedings of the workshop on noisy user-generated text, p. 126–135. Cited by: §A.1, §A.1, §A.2, §2.1, §2.3. Bisani and Ney (2008) M. Bisani and H. Ney Joint-sequence models for grapheme-to-phoneme conversion. Speech communication 50 (5), p. 434–451. Cited by: §1, §2.1. Brody and Diakopoulos (2011) S. Brody and N. Diakopoulos Cooooooooooooooollllllllllllll!!!!!!!!!!!!!! using word lengthening to detect sentiment in microblogs. In Proceedings of the 2011 conference on empirical methods in natural language processing, p. 562–570. Cited by: §2.2. Buaphet et al. (2026) W. Buaphet, T. Nguyen, R. Kondo, T. Kajiwara, Y. Kim, J. Lee, H. Lee, H. Lovenia, P. Limkonchotiwat, S. Nutanong, and R. Van der Goot MultiLexNorm++: a unified benchmark and a generative model for lexical normalization for asian languages. ACM Transactions on Asian and Low-Resource Language Information Processing. Cited by: §A.1, §A.1, §2.1, §2.3. Choi et al. (2023) D. Choi, J. Song, E. Lee, J. Seo, H. Park, and D. Na KoMultiText: large-scale korean text dataset for classifying biased speech in real-world online services. arXiv preprint arXiv:2310.04313. Cited by: §A.1, §A.1, §2.1, §2.3. Doherty (2016) L. Doherty IPA dict: monolingual wordlists with pronunciation information in ipa. Note: https://github.com/open-dict-data/ipa-dictAccessed 2026-05-22 Cited by: §A.4, §1, §2.3, §4.1. Dong et al. (2022) L. Dong, Z. Guo, C. Tan, Y. Hu, Y. Jiang, and Z. Ling Neural grapheme-to-phoneme conversion with pre-trained grapheme models. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 6202–6206. Cited by: Appendix D, Table 3, §4.1. Han and Baldwin (2011) B. Han and T. Baldwin Lexical normalisation of short text messages: makn sens a# twitter. In Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies, p. 368–378. Cited by: §1. Han et al. (2024) D. Han, M. Cui, J. Kang, X. Wu, X. Liu, and H. Meng Improving grapheme-to-phoneme conversion through in-context knowledge retrieval with large language models. In 2024 IEEE 14th International Symposium on Chinese Spoken Language Processing (ISCSLP), p. 631–635. Cited by: Appendix D, §4.1. Kim et al. (2021) J. Kim, J. Kong, and J. Son Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In International conference on machine learning, p. 5530–5540. Cited by: §1. Landis and Koch (1977) J. R. Landis and G. G. Koch The measurement of observer agreement for categorical data. biometrics, p. 159–174. Cited by: §A.4, §2.3. Lee and Ashok (2022) H. Lee and V. Ashok Impact of out-of-vocabulary words on the twitter experience of blind users. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, p. 1–20. Cited by: §1. Lee et al. (2020) J. L. Lee, L. F. Ashby, M. E. Garza, Y. Lee-Sikka, S. Miller, A. Wong, A. D. McCarthy, and K. Gorman Massively multilingual pronunciation modeling with wikipron. In Proceedings of the twelfth language resources and evaluation conference, p. 4223–4228. Cited by: §1. Mortensen et al. (2018) D. R. Mortensen, S. Dalmia, and P. Littell Epitran: precision g2p for many languages. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Cited by: Appendix D, §4.1, §4.2. Nguyen et al. (2024) T. Nguyen, T. Le, and K. Van Nguyen ViLexNorm: a lexical normalization corpus for vietnamese social media text. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), p. 1421–1437. Cited by: §A.1, §A.1, §A.2, §2.1, §2.3. Novak et al. (2012) J. R. Novak, N. Minematsu, and K. Hirose WFST-based grapheme-to-phoneme conversion: open source tools for alignment, model-building and decoding. In Proceedings of the 10th International Workshop on Finite State Methods and Natural Language Processing, p. 45–49. Cited by: §1, §2.1. Plank (2016) B. Plank What to do about non-standard (or non-canonical) language in nlp. In Conference on Natural Language Processing (KONVENS), Cited by: §1. Ploujnikov and Ravanelli (2022) A. Ploujnikov and M. Ravanelli SoundChoice: grapheme-to-phoneme models with semantic disambiguation. In Interspeech, Cited by: §1. Qwen et al. (2025) Qwen, :, A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Tang, T. Xia, X. Ren, X. Ren, Y. Fan, Y. Su, Y. Zhang, Y. Wan, Y. Liu, Z. Cui, Z. Zhang, and Z. Qiu Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. Cited by: Appendix D, §3.1. Radford et al. (2023) A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever Robust speech recognition via large-scale weak supervision. In International conference on machine learning, p. 28492–28518. Cited by: Appendix F, §4.3. Rao et al. (2015) K. Rao, F. Peng, H. Sak, and F. Beaufays Grapheme-to-phoneme conversion using long short-term memory recurrent neural networks. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 4225–4229. Cited by: §1, §2.1. Ren et al. (2019) Y. Ren, Y. Ruan, X. Tan, T. Qin, S. Zhao, Z. Zhao, and T. Liu Fastspeech: fast, robust and controllable text to speech. Advances in neural information processing systems 32. Cited by: §1. Shen et al. (2018) J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y. Zhang, Y. Wang, R. Skerrv-Ryan, R. A. Saurous, Y. Agiomyrgiannakis, and Y. Wu Natural tts synthesis by conditioning wavenet on mel spectrogram predictions. In 2018 IEEE international conference on acoustics, speech and signal processing (ICASSP), p. 4779–4783. Cited by: §1. Singh et al. (2025) A. Singh, A. Fry, A. Perelman, A. Tart, A. Ganesh, A. El-Kishky, A. McLaughlin, A. Low, A. Ostrow, A. Ananthram, A. Nathan, A. Luo, A. Helyar, A. Madry, A. Efremov, A. Spyra, A. Baker-Whitcomb, A. Beutel, A. Karpenko, A. Makelov, A. Neitz, A. Wei, A. Barr, A. Kirchmeyer, A. Ivanov, A. Christakis, A. Gillespie, A. Tam, A. Bennett, A. Wan, A. Huang, A. M. Sandjideh, A. Yang, A. Kumar, A. Saraiva, A. Vallone, A. Gheorghe, A. G. Garcia, A. Braunstein, A. Liu, A. Schmidt, A. Mereskin, A. Mishchenko, A. Applebaum, A. Rogerson, A. Rajan, A. Wei, A. Kotha, A. Srivastava, A. Agrawal, A. Vijayvergiya, A. Tyra, A. Nair, A. Nayak, B. Eggers, B. Ji, B. Hoover, B. Chen, B. Chen, B. Barak, B. Minaiev, B. Hao, B. Baker, B. Lightcap, B. McKinzie, B. Wang, B. Quinn, B. Fioca, B. Hsu, B. Yang, B. Yu, B. Zhang, B. Brenner, C. R. Zetino, C. Raymond, C. Lugaresi, C. Paz, C. Hudson, C. Whitney, C. Li, C. Chen, C. Cole, C. Voss, C. Ding, C. Shen, C. Huang, C. Colby, C. Hallacy, C. Koch, C. Lu, C. Kaplan, C. Kim, C. Minott-Henriques, C. Frey, C. Yu, C. Czarnecki, C. Reid, C. Wei, C. Decareaux, C. Scheau, C. Zhang, C. Forbes, D. Tang, D. Goldberg, D. Roberts, D. Palmie, D. Kappler, D. Levine, D. Wright, D. Leo, D. Lin, D. Robinson, D. Grabb, D. Chen, D. Lim, D. Salama, D. Bhattacharjee, D. Tsipras, D. Li, D. Yu, D. Strouse, D. Williams, D. Hunn, E. Bayes, E. Arbus, E. Akyurek, E. Y. Le, E. Widmann, E. Yani, E. Proehl, E. Sert, E. Cheung, E. Schwartz, E. Han, E. Jiang, E. Mitchell, E. Sigler, E. Wallace, E. Ritter, E. Kavanaugh, E. Mays, E. Nikishin, F. Li, F. P. Such, F. d. A. B. Peres, F. Raso, F. Bekerman, F. Tsimpourlas, F. Chantzis, F. Song, F. Zhang, G. Raila, G. McGrath, G. Briggs, G. Yang, G. Parascandolo, G. Chabot, G. Kim, G. Zhao, G. Valiant, G. Leclerc, H. Salman, H. Wang, H. Sheng, H. Jiang, H. Wang, H. Jin, H. Sikchi, H. Schmidt, H. Aspegren, H. Chen, H. Qiu, H. Lightman, I. Covert, I. Kivlichan, I. Silber, I. Sohl, I. Hammoud, I. Clavera, I. Lan, I. Akkaya, I. Kostrikov, I. Kofman, I. Etinger, I. Singal, J. Hehir, J. Huh, J. Pan, J. Wilczynski, J. Pachocki, J. Lee, J. Quinn, J. Kiros, J. Kalra, J. Samaroo, J. Wang, J. Wolfe, J. Chen, J. Wang, J. Harb, J. Han, J. Wang, J. Zhao, J. Chen, J. Yang, J. Tworek, J. Chand, J. Landon, J. Liang, J. Lin, J. Liu, J. Wang, J. Tang, J. Yin, J. Jang, J. Morris, J. Flynn, J. Ferstad, J. Heidecke, J. Fishbein, J. Hallman, J. Grant, J. Chien, J. Gordon, J. Park, J. Liss, J. Kraaijeveld, J. Guay, J. Mo, J. Lawson, J. McGrath, J. Vendrow, J. Jiao, J. Lee, J. Steele, J. Wang, J. Mao, K. Chen, K. Hayashi, K. Xiao, K. Salahi, K. Wu, K. Sekhri, K. Sharma, K. Singhal, K. Li, K. Nguyen, K. Gu-Lemberg, K. King, K. Liu, K. Stone, K. Yu, K. Ying, K. Georgiev, K. Lim, K. Tirumala, K. Miller, L. Ahmad, L. Lv, L. Clare, L. Fauconnet, L. Itow, L. Yang, L. Romaniuk, L. Anise, L. Byron, L. Pathak, L. Maksin, L. Lo, L. Ho, L. Jing, L. Wu, L. Xiong, L. Mamitsuka, L. Yang, L. McCallum, L. Held, L. Bourgeois, L. Engstrom, L. Kuhn, L. Feuvrier, L. Zhang, L. Switzer, L. Kondraciuk, L. Kaiser, M. Joglekar, M. Singh, M. Shah, M. Stratta, M. Williams, M. Chen, M. Sun, M. Cayton, M. Li, M. Zhang, M. Aljubeh, M. Nichols, M. Haines, M. Schwarzer, M. Gupta, M. Shah, M. Y. Guan, M. Huang, M. Dong, M. Wang, M. Glaese, M. Carroll, M. Lampe, M. Malek, M. Sharman, M. Zhang, M. Wang, M. Pokrass, M. Florian, M. Pavlov, M. Wang, M. Chen, M. Wang, M. Feng, M. Bavarian, M. Lin, M. Abdool, M. Rohaninejad, N. Soto, N. Staudacher, N. LaFontaine, N. Marwell, N. Liu, N. Preston, N. Turley, N. Ansman, N. Blades, N. Pancha, N. Mikhaylin, N. Felix, N. Handa, N. Rai, N. Keskar, N. Brown, O. Nachum, O. Boiko, O. Murk, O. Watkins, O. Gleeson, P. Mishkin, P. Lesiewicz, P. Baltescu, P. Belov, P. Zhokhov, P. Pronin, P. Guo, P. Thacker, Q. Liu, Q. Yuan, Q. Liu, R. Dias, R. Puckett, R. Arora, R. T. Mullapudi, R. Gaon, R. Miyara, R. Song, R. Aggarwal, R. Marsan, R. Yemiru, R. Xiong, R. Kshirsagar, R. Nuttall, R. Tsiupa, R. Eldan, R. Wang, R. James, R. Ziv, R. Shu, R. Nigmatullin, S. Jain, S. Talaie, S. Altman, S. Arnesen, S. Toizer, S. Toyer, S. Miserendino, S. Agarwal, S. Yoo, S. Heon, S. Ethersmith, S. Grove, S. Taylor, S. Bubeck, S. Banesiu, S. Amdo, S. Zhao, S. Wu, S. Santurkar, S. Zhao, S. R. Chaudhuri, S. Krishnaswamy, S. Xia, S. Cheng, S. Anadkat, S. P. Fishman, S. Tobin, S. Fu, S. Jain, S. Mei, S. Egoian, S. Kim, S. Golden, S. Mah, S. Lin, S. Imm, S. Sharpe, S. Yadlowsky, S. Choudhry, S. Eum, S. Sanjeev, T. Khan, T. Stramer, T. Wang, T. Xin, T. Gogineni, T. Christianson, T. Sanders, T. Patwardhan, T. Degry, T. Shadwell, T. Fu, T. Gao, T. Garipov, T. Sriskandarajah, T. Sherbakov, T. Korbak, T. Kaftan, T. Hiratsuka, T. Wang, T. Song, T. Zhao, T. Peterson, V. Kharitonov, V. Chernova, V. Kosaraju, V. Kuo, V. Pong, V. Verma, V. Petrov, W. Jiang, W. Zhang, W. Zhou, W. Xie, W. Zhan, W. McCabe, W. DePue, W. Ellsworth, W. Bain, W. Thompson, X. Chen, X. Qi, X. Xiang, X. Shi, Y. Dubois, Y. Yu, Y. Khakbaz, Y. Wu, Y. Qian, Y. T. Lee, Y. Chen, Y. Zhang, Y. Xiong, Y. Tian, Y. Cha, Y. Bai, Y. Yang, Y. Yuan, Y. Li, Y. Zhang, Y. Yang, Y. Jin, Y. Jiang, Y. Wang, Y. Wang, Y. Liu, Z. Stubenvoll, Z. Dou, Z. Wu, and Z. Wang Openai gpt-5 system card. arXiv preprint arXiv:2601.03267. Cited by: §C.1, Appendix D, §3.2, §4.1, §4.3. Taylor (2005) P. Taylor Hidden markov models for grapheme to phoneme conversion.. In Interspeech, p. 1973–1976. Cited by: §1. Tri et al. (2020) N. M. Tri C. X. Nam et al. Vietnamese speech synthesis with end-to-end model and text normalization. In 2020 7th NAFOSTED Conference on Information and Computer Science (NICS), p. 179–184. Cited by: Appendix D, §4.1. Van Der Goot et al. (2021) R. Van Der Goot, A. Ramponi, A. Zubiaga, B. Plank, B. Muller, I. S. V. Roncal, N. Ljubešić, Ö. Çetinoğlu, R. Mahendra, T. Çolakoğlu, T. Baldwin, T. Caselli, and W. Sidorenko MultiLexNorm: a shared task on multilingual lexical normalization. In Proceedings of the Seventh Workshop on Noisy User-generated Text (W-NUT 2021), p. 493–509. Cited by: §A.1, §A.1, §2.1, §2.3. Van Der Goot et al. (2018) R. Van Der Goot, R. Van Noord, and G. van Noord A taxonomy for in-depth evaluation of normalization for user generated content. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Cited by: §2.1, §2.2, §2.2. Van Der Goot (2019) R. Van Der Goot MoNoise: a multi-lingual and easy-to-use lexical normalization tool. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, p. 201–206. Cited by: Appendix D, §4.1, §4.3, Table 5. Van Esch and Sproat (2017) D. Van Esch and R. Sproat An expanded taxonomy of semiotic classes for text normalization.. In Interspeech, p. 4016–4020. Cited by: §1. Wong et al. (2025) M. Wong, A. Alshehri, S. Kao, and H. He PolyNorm: few-shot llm-based text normalization for text-to-speech. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, p. 77–85. Cited by: Appendix D, §4.1. Xue et al. (2022) L. Xue, A. Barua, N. Constant, R. Al-Rfou, S. Narang, M. Kale, A. Roberts, and C. Raffel ByT5: towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics 10, p. 291–306. Cited by: Appendix D, Appendix D, §3.1, Table 3. Yolchuyeva et al. (2019) S. Yolchuyeva, G. Németh, and B. Gyires-Tóth Transformer based grapheme-to-phoneme conversion. In Interspeech, Cited by: Appendix D, §1, Table 3, §4.1, §4.2. Yoon et al. (2023) E. Yoon, H. S. Yoon, D. Gowda, S. Eom, D. Kim, J. Harvill, H. Gao, M. Hasegawa-Johnson, C. Kim, and C. D. Yoo Mitigating the exposure bias in sentence-level grapheme-to-phoneme (g2p) transduction. In Interspeech, Cited by: §3.1, §4.3. Zhao et al. (2022) C. Zhao, J. Wang, X. Qu, H. Wang, and J. Xiao R-g2p: evaluating and enhancing robustness of grapheme to phoneme conversion by controlled noise introducing and contextual information incorporation. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 6197–6201. Cited by: Appendix D, §1, §4.1. Zhu et al. (2022) J. Zhu, C. Zhang, and D. Jurgens ByT5 model for massively multilingual grapheme-to-phoneme conversion. In Interspeech, Cited by: §1, Table 3, §4.1, §4.2. Appendix A Creating UGTPhon This appendix details the dataset construction procedure summarized in Section 2.3, including the procedures for source corpus collection and preprocessing, as well as the annotation pipelines for canonical forms, categories, and phonemes. A.1 Source Corpora and Preprocessing Source corpora. Non-canonical tokens GncG_nc in UGTPhon are drawn from existing user-generated text resources, with small human-curated augmented subsets for English and Vietnamese. For English, we use MultiLexNorm (Van Der Goot et al., 2021), whose English data builds on LexNorm2015 (Baldwin et al., 2015). For Vietnamese, we use ViLexNorm (Nguyen et al., 2024), a lexical normalization corpus of Vietnamese social media text whose annotations align with our Gnc→GcG_nc→ G_c formulation. For Korean, we source GncG_nc from MultiLexNorm++ (Buaphet et al., 2026), supplemented by the Korean UGT corpus of Choi et al. (2023), and re-annotate GcG_c from scratch using the protocol in Section A.2. In addition, the English and Vietnamese subsets include 175 and 385 human-curated augmented UGT samples, respectively, generated from existing UGT patterns. Licenses. Source corpora are used under their original licensing terms. MultiLexNorm’s English data builds on LexNorm2015 (Baldwin et al., 2015), which has no explicit license; following the convention adopted by MultiLexNorm (Van Der Goot et al., 2021), attribution is provided through citation. The English subset of UGTPhon is provided strictly for academic and research purposes, and commercial use is prohibited. ViLexNorm (Nguyen et al., 2024) is released under C BY-NC-SA 4.0. The Korean subset is derived from KoMultiText (Choi et al., 2023), partly via MultiLexNorm++ (Buaphet et al., 2026), and is released under Apache 2.0. UGTPhon is released with per-language licenses reflecting these upstream terms. Data provenance and consent. UGTPhon is constructed from previously released UGT datasets rather than through direct collection from the original social-media users. We therefore do not separately recruit or obtain consent from the original text authors, and instead follow the data-use conditions and licensing terms of the respective source datasets. Preprocessing. We preserve the original sentence boundaries, casing, and surrounding context of each GncG_nc token in the UGT sentence. Non-lexical elements are handled during phoneme construction rather than uniformly removed from the raw text: punctuation may be dropped, numbers may be verbalized, and elements that cannot be phonemized may remain as placeholders. A.2 Canonical Form Annotation For English and Vietnamese, canonical forms GcG_c are obtained directly from the source lexical normalization corpora (Baldwin et al., 2015; Nguyen et al., 2024), which provide surface-to-canonical mappings. Manual annotation for Korean. Unlike English and Vietnamese, no comparably large-scale lexical normalization corpus exists for Korean. We therefore annotate GcG_c from scratch with two native-speaker experts following a category-specific protocol that reflects the structure of Korean UGT; the full annotation guide is provided below: • A1 (Shortening-V, 초성체): Consonant abbreviations are expanded to their full Hangul form (e.g., ㄱㅅ→ 감사, ㅇㄱㄹㅇ→ 이거레알). • B1 (Lengthening): The elongation of vowel or syllable is collapsed to their base form (e.g., 꺄아악→ 꺅). • B2 (Repetition): For consonant-only repetition (e.g., ㅋ), GcG_c is the Hangul realization preserving the intended count (e.g., 크, 키, 킥; recorded as comma-separated alternatives when more than one is acceptable). For syllable- or word-level repetition (e.g., 빨리빨리, 감사감사), the surface itself is the intended pronunciation, so GcG_c retains the repetition. • C1–C4 (Pass-through): The surface form already encodes the intended pronunciation, so Gc=GncG_c=G_nc (e.g., 머해, 킹받다, 갑분싸). Tokens from the source corpora that turn out to be standard Korean rather than UGT are removed from the benchmark. A.3 Category Annotation Type ID Category Description Examples A Reconstruct A1 Shortening-V (Vowel Drop) Canonical-form consonants are preserved, vowels are dropped. pls → please đc → được A2 Shortening-O (Other Drop) One or more characters dropped from a single canonical word, beyond vowel-only deletion. bc → because k → không A3 Phrasal (Unpronounceable) A multi-word canonical expansion whose pronunciation is not directly recoverable from the surface form. idk → I don’t know mn → mọi người B Repetition B1 Lengthening Repeated characters convey prosodic emphasis; the canonical form collapses the repetition. soooo → so B2 Iteration Repeated tokens or words; the canonical form retains the repetition for phonemization. wait wait C Pass-through C1 Eye Direct A phonetic respelling whose surface form already encodes the intended pronunciation. luv → love zô → vô C2 Regular Standard productive variation of a canonical word. droppin → dropping C3 Slang/Teen Lexicalized non-standard form whose pronunciation is well established. tho → though hong → không C4 Phrasal (Pronounceable) Multi-letter abbreviation pronounced as a word in its own right. rofl GATO Table 7: Operational definitions and representative examples of the fine-grained categories for non-canonical words. Category definitions. We use the nine fine-grained categories listed in Table 1, grouped into the three macro-types, Reconstruct (A1–A3), Repetition (B1–B2), and Pass-through (C1–C4). The macro-grouping reflects the type of inference required to phonemize a token, rather than its surface transformation. Table 7 presents the operational definitions and representative examples used during annotation. Category annotation pipeline. Category labels are assigned manually by native-speaker annotators following the operational definitions in Table 7. For each Gnc→GcG_nc _c pair, an annotator inspects the surrounding sentence and selects the single best-fitting category among A1–C4. Each pair is then independently verified by a second native-speaker annotator, with disagreements resolved through discussion. All annotators have prior training in NLP or linguistics and are instructed to flag examples where (i) the canonical form is genuinely ambiguous given the context, or (i) the category assignment is unclear; cases of type (i) are resolved using sentence context as described in Section 2.1, and cases of type (i) default to the more specific category (e.g., A3 rather than A1 when both apply). A.4 Phoneme Annotation Target phonemes PcP_c are derived from the canonical form GcG_c using IPA-Dict (Doherty, 2016). For English heteronyms (e.g., lead as /li:d/ or /lEd/), the correct pronunciation depends on sentence-level context and IPA-Dict alone is insufficient. Therefore, an annotator with linguistic expertise manually selects the appropriate phoneme sequence based on the surrounding sentence. Out-of-vocabulary handling. Canonical forms not covered by IPA-Dict are manually transcribed following the same IPA conventions. We measure OOV over unique canonical-form types across all splits, using the corresponding IPA-Dict lexicons (en_US, vi_N, and ko); non-phonemizable and purely symbolic forms are excluded. Across all canonical types, the OOV rates are 39.4% (EN), 23.1% (VI), and 77.6% (KO). Restricted to non-canonical UGT targets, they are 4.3% (48/1,110), 7.4% (175/2,378), and 92.8% (1,991/2,146), respectively. The high Korean rate largely reflects the mismatch between IPA-Dict entries and the inflected canonical forms preserved in UGTPhon. Language Tokens Cohen’s κ Raw Agreement by UGT Type (%) Overall A1 A2 A3 B1 B2 C1 C2 C3 C4 English 1,327 0.860 86.3 98.2 87.8 51.3 100.0 86.4 92.1 88.1 88.6 67.7 Vietnamese 2,152 0.869 87.3 84.1 92.7 83.6 97.7 48.6 89.4 92.3 91.6 0.0 Korean 722 0.754 75.6 87.9 — — 78.8 68.1 86.4 90.9 70.9 62.5 Table 8: Inter-annotator agreement for phoneme annotations across languages and UGT types. While English and Vietnamese show near-perfect agreement overall, Korean exhibits lower agreement due to greater variation in IPA transcription conventions. Inter-annotator agreement. To assess phoneme annotation reliability, two annotators independently perform the procedure above, and we compute agreement between the resulting transcriptions. As shown in Table 8, Cohen’s κ over IPA phoneme sequences reaches 0.86 for English and 0.87 for Vietnamese, with Korean at 0.75. The Korean score, while lower, remains within the substantial band (Landis and Koch, 1977). This mainly reflects divergent IPA transcription conventions, particularly differing choices of consonant allophone symbols. We report the detailed agreement rates by UGT type in Table 8. For English, agreement remains consistently high across most categories, although relatively lower agreement was observed for A3 and C4. Vietnamese shows a similar overall pattern, with lower agreement for B2 and especially C4. Notably, the zero agreement for Vietnamese C4 primarily reflects genuine ambiguity in letter-by-letter IPA renderings of acronyms rather than annotation errors. Overall, these results indicate that the annotation process produces reliable phoneme transcriptions across languages while also revealing categories where pronunciation conventions are inherently more variable. For the released dataset, we resolve all inter-annotator conflicts through discussion between the two annotators to produce a single adjudicated version. Appendix B Content Moderation Analysis Language Sentences Flagged (raw) Flagged (norm) English 5,029 568 (11.29%) 606 (12.05%) Vietnamese 10,847 401 (3.70%) 473 (4.36%) Korean 2,497 321 (12.86%) 322 (12.90%) Total 18,373 1,290 (7.02%) 1,401 (7.63%) Table 9: Sentence-level moderation flagged rates on UGTPhon, evaluated with OpenAI omni-moderation-2024-09-26 on both raw UGT forms and canonical-form normalizations. A sentence is flagged if any of the 13 moderation categories triggers. User-generated text often contains harmful or offensive content, and UGTPhon inherits this property from its source corpora. To characterize the toxicity profile of UGTPhon and to measure how canonical-form normalization affects moderation signal, we apply OpenAI’s omni-moderation-2024-09-26 model to every sentence in the corpus, scoring both the raw UGT form and its canonical-form normalization across 1313 moderation categories. B.1 Flagged Rates Table 9 reports sentence-level flagged rates. Korean shows the highest flagged rate (12.9%12.9\%), comparable to English (11.3%11.3\%) and roughly three times higher than Vietnamese (3.7%3.7\%). The dominant category is harassment in English and Korean, and violence in Vietnamese, reflecting source-corpus content rather than language-intrinsic properties. B.2 Effect of Normalization In general, canonical-form normalization marginally increases the overall flagged rate. The binary flag state changes for 385 of 18,373 sentences (2.10%), while the aggregate flagged rate increases by only 0.60 percentage points (7.02% → 7.63%). This indicates that most sentences retain the same moderation outcome under normalization. The increase is most pronounced in Vietnamese (3.70% → 4.36%), while English (11.29% → 12.05%) and Korean (12.86% → 12.90%) show smaller changes. Appendix C LLM-Based Datastore Expansion This appendix details the LLM-based expansion of the retrieval datastore D summarized in Section 3.2. C.1 Prompt Template We prompt GPT-5.4-mini Singh et al. (2025) with seed examples from the training set to generate additional plausible (Gnc,Gc)(G_nc,G_c) pairs. The prompt is structured as follows: (i) a short description of UGT and the desired output format, (i) k seed pairs sampled from the training datastore, and (i) an instruction to generate n additional pairs that are similar in style but not identical to the seeds. We use k seed pairs sampled from all training entries of the given UGT type (ranging from a few dozen to over 200 per type depending on training-set density) and set n∈150,100,200n∈\150,100,200\ for types A1, A2, and A3, respectively. The remaining categories (B1, B2, C1–C4) are populated from the training-set entries only, without LLM-generated augmentation. The full prompt template is given below: nc_type, type_desc, lang_name, seed_lines, and target are filled per type. The system message instructs the model to act as an expert in the target language’s internet slang. Here are real examples of nc_type type UGT tokens from lang_name social media data. nc_type definition: type_desc Existing examples (do NOT repeat these): seed_lines Generate target MORE nc_type tokens commonly used in lang_name Twitter/Instagram/Discord/Reddit/texting. Requirements: - Follow the same pattern as nc_type definition above. - Each expansion must be the standard lang_name form. - Do NOT include any of the examples above. - Do NOT include standard dictionary words as the raw form. - Do NOT include readable acronyms that are pronounced letter by letter. - Cover a wide variety: different topics, registers, formality levels. - Return ONLY a JSON object: "raw_token": "expanded form", ... C.2 Filtering Criteria Each generated (Gnc,Gc)(G_nc,G_c) candidate is filtered before insertion into D. We discard candidates that satisfy any of the following conditions: • Trivial identity: Gc=GncG_c=G_nc, meaning no normalization is needed and inserting the pair would provide no useful retrieval signal. • Standard token conflict: GncG_nc matches a canonical token already present in the training or development set, which would risk over-correcting canonical inputs at inference. • Test set leakage: The test set is never inspected during generation or filtering. Candidates whose GncG_nc happens to coincide with a test non-canonical token are retained only if the same surface form already appears in the training set with a consistent GcG_c. C.3 Coverage Statistics The final datastore D used in our main results combines training-set entries with LLM-augmented pairs. The training-only datastore contains 1,413/3,674/1,6071,413/3,674/1,607 unique non-canonical surface-form keys for English, Vietnamese, and Korean, respectively. The LLM-based expansion adds 491/220/246491/220/246 entries, respectively. We define test-set coverage as the proportion of non-canonical target-token occurrences whose surface form GncG_nc has an exact-match key in the datastore. Under this definition, the training-only datastore covers 70.9%70.9\% (941/1,327941/1,327), 82.9%82.9\% (1,785/2,1521,785/2,152), and 25.3%25.3\% (183/722183/722) of the English, Vietnamese, and Korean test targets, respectively. Appendix D Training and Implementation Details Backbones. We test our approach with two transformer backbones: ByT5-small (Xue et al., 2022) (300M parameters, byte-level encoder-decoder, default) and Qwen2.5-0.5B (Qwen et al., 2025) (500M parameters, decoder-only causal LM). Optimization. All backbones are trained with AdamW (learning rate ×10−41\!×\!10^-4, batch size 3232, 1010 epochs) on 4 NVIDIA V100 GPUs. We select the checkpoint with the lowest development set PERncPER_nc. Compute budget. Each of our method’s training runs takes approximately 66 hours on 44 NVIDIA V100 GPUs. Including baseline retraining, backbone ablation, and frontier-LLM inference, the total compute used for all results reported in the paper is approximately 400400 V100-GPU hours. Frontier-LLM API calls are not included in this figure. Decoding. During inference, the model decodes greedily. For non-canonical tokens, the model first emits GcG_c following the <Target> marker, then the phoneme sequence PcP_c following <Phoneme>; for canonical tokens, PcP_c is emitted directly without the <Target> step. Baselines. The rule-based baselines are Epitran (Mortensen et al., 2018) and viG2P (Tri et al., 2020), both used out-of-the-box with default settings. For viG2P, we use the Viphoneme implementation.22 2 https://github.com/v-nhandt21/Viphoneme The supervised neural baselines comprise Transformer (Yolchuyeva et al., 2019), ByT5 and ByT5-tiny (Xue et al., 2022), and GBERT (Dong et al., 2022); each is trained on the IPA-Dict training split with the same hyperparameters as in its original paper. The frontier-LLM baselines are GPT-4o, GPT-5.4-mini (Singh et al., 2025), Gemini-2.5-flash, Gemini-3-flash, and Claude-Sonnet-4.6, queried via their official APIs (model identifiers: gpt-4o, gpt-5.4-mini, gemini-2.5-flash, gemini-3-flash, and claude-sonnet-4-6; all accessed May 2026). For the K=9K=9 setting, we use one in-context example per available UGT category, yielding nine examples for English and Vietnamese and seven for Korean, where A2 and A3 are absent. For the K∈0,9,27K∈\0,9,27\ shot-scaling study, this corresponds to 37,440 frontier-LLM requests (2,496 test sentences × 5 models × 3 shot settings). The task-relevant baselines are reproduced or adapted following their published descriptions and evaluated with the same UGTPhon splits and PER pipeline: ICKR (Han et al., 2024) uses GPT-4o with in-context knowledge retrieval; R-G2P (Zhao et al., 2022) uses its controlled-noise robustness training; MoNoise (Van Der Goot, 2019) (English only) and PolyNorm (Wong et al., 2025) perform lexical normalization before the same word-level ByT5 G2P backend. English Vietnamese Korean Model PERcPER_c PERncPER_nc PERcPER_c PERncPER_nc PERcPER_c PERncPER_nc ByT5 (300M) 5.2 29.8 7.7 43.2 4.3 47.5 Qwen2.5-0.5B (naive FT) 11.2 14.8 4.4 23.1 14.2 77.3 ICKR (GPT-4o, reproduced) 7.7 14.9 7.9 30.8 8.2 47.9 R-G2P (reproduced) 18.2 53.2 6.9 59.3 4.1 42.3 PolyNorm (GPT-4) + ByT5 25.8 44.5 15.6 49.3 15.5 28.5 MoNoise + ByT5 14.2 18.0 – – – – Ours (Qwen2.5-0.5B) 10.2 14.3 4.0 20.0 6.0 32.4 Ours (ByT5, 300M) 4.7 22.5 5.9 26.7 5.1 32.0 Table 10: Task-relevant and matched-backbone comparisons on UGTPhon (PER, %). Bold marks the best value among the listed systems for each metric. Normalization-first systems can improve PERncPER_nc while substantially increasing PERcPER_c. Multi-seed stability. We rerun both proposed variants with three random seeds and report mean ± sample standard deviation in Table 11. English and Vietnamese are stable; Korean varies more, consistent with its smaller test split and larger unseen-form rate. Backbone Lang. PERcPER_c PERncPER_nc Qwen2.5-0.5B EN ±0.1510.07\!±\!0.15 ±0.3114.03\!±\!0.31 VI ±0.063.93\!±\!0.06 ±0.0620.03\!±\!0.06 KO ±1.394.47\!±\!1.39 ±2.6029.57\!±\!2.60 ByT5 EN ±0.104.70\!±\!0.10 ±2.0720.60\!±\!2.07 VI ±0.065.83\!±\!0.06 ±0.3227.07\!±\!0.32 KO ±0.324.73\!±\!0.32 ±1.5930.17\!±\!1.59 Table 11: Three-seed mean ± sample SD PER (%). Paired statistical tests. We additionally run two-sided paired bootstrap tests with 10,000 sentence-level resamples on the checkpoints reported in the main tables. Table 12 reports Δ=Ours−matched baseline =Ours-matched baseline for PERncPER_nc. All intervals exclude zero, although the English Qwen effect is small. Lang. Backbone Ours Base 95% CI of Δ EN Qwen 14.3 14.8 [−0.8,−0.1][-0.8,-0.1] EN ByT5 22.5 29.8 [−8.9,−3.4][-8.9,-3.4] VI Qwen 20.0 23.1 [−5.4,−1.3][-5.4,-1.3] VI ByT5 26.7 43.2 [−18.3,−13.8][-18.3,-13.8] KO Qwen 32.4 77.3 [−47.9,−43.1][-47.9,-43.1] KO ByT5 32.0 47.5 [−18.4,−10.3][-18.4,-10.3] Table 12: Paired-bootstrap PERncPER_nc comparisons (10,000 sentence-level resamples). Cat. User (token list) Assistant (IPA list) A1 Some / guys / hv / a / bad / game ”s@m / ”gaIz / ”hæv / @ / ”bæd / ”geIm A2 bruh / y / u / not / playin br2 / ”waI / ”ju / ”nAt / ”pleIIŋ A3 naah / im / just / idk / sorry ”nA / ”aIm / ”dZ@st / ”aI doUnt noU / ”sAri B1 yeah / like / a / loooot ”jæ/ ”laIk / ”eI / ”lOt B2 frfr / goodmorning fr fr / ”gUd”mOrnIŋ C1 wat / u / kno / bout / hockey ”w@t / ”ju / ”noU / @”baUt / ”hAki C2 Wats / goin / on ”hw@ts / ”goUIŋ/ ”An C3 goin / wit / da / crew ”goUIŋ/ ”wIT / D@ / ”kru C4 i / thought / of / u / omg ”aI / ”TOt / @v / ”ju / ”oU ”maI ”gAd Table 13: English few-shot example pool: one representative (user, assistant) pair per UGT category, drawn first when K=9K=9. “/” separates the per-token entries that the prompt presents on numbered lines. Appendix E Frontier LLM Prompt Template This section describes the in-context learning setup used to query the frontier-LLM baselines (Appendix D). We use the same chat-message structure across GPT-4o, GPT-5.4-mini, Gemini-2.5-flash, Gemini-3-flash, and Claude-Sonnet-4.6, with decoding parameters temperature=0 and max_completion_tokens=512. Message structure. For each input sentence S, we issue a single chat request consisting of (i) a language-specific system message, (i) a set of few-shot exchanges, each as a (user, assistant) pair, and (i) the query sentence as the final user turn: [system] language-specific G2P instructions [user] 1. tok1 2. tok2 … [assistant] 1. ipa1 2. ipa2 … … (few-shot pairs) [user] 1. w1w_1 2. w2w_2 … Both input and output use a numbered one-token-per-line format, and responses are parsed with the regex ^( +)[.): ] *(.*)$. Each token in a sentence is presented to the model as one numbered line, and the model returns the IPA transcription on the same line number. Shot scaling. The nominal K∈0,9,27K∈\0,9,27\ settings in Figure 4 correspond to: K=0K=0, with no few-shot pairs; K=9K=9, with one pair per available UGT category; and K=27K=27, with three pairs per available UGT category. Thus, the actual numbers of in-context examples are 0/9/270/9/27 for English and Vietnamese and 0/7/210/7/21 for Korean, since A2 and A3 do not appear in Korean. Few-shot example pairs are sampled from the training split, drawn once per UGT category, and reused across all queries. System messages. The system message specifies the G2P task and target IPA conventions for each language. We reproduce the English system message below; the Vietnamese and Korean variants follow the same template, written in the target language and referencing language-specific phonological conventions (Vietnamese tone marks, Korean tense consonants and consonant assimilation). You are a grapheme-to-phoneme (G2P) expert for English. You will receive a numbered list of tokens from a sentence (possibly containing informal, abbreviated, or slang forms). Output the IPA transcription for each token in the same numbered order. Rules: - Each line of input is ONE token, identified by its number (e.g. ‘3. idk’). - Output exactly one numbered IPA line per token: ‘3. "aI doUnt noU’. - A token may contain spaces (e.g. a NeMo-expanded handle) --- treat it as one unit. - For abbreviated/informal tokens (e.g. pls→ , idk→ don’t know, sooo→ ), first determine the standard form, then transcribe. - For repeated chars/words (e.g. frfr, sooooo), transcribe the normalized form. - Use standard IPA notation. - Output ONLY numbered IPA lines, no explanations or extra text. Few-shot example pools. For each language and UGT category, we maintain a fixed pool of three (token list, IPA list) pairs and draw n pairs per available category in order, where n=1n=1 for the nominal K=9K=9 setting and n=3n=3 for the nominal K=27K=27 setting. Because Korean lacks A2 and A3, these settings contain 7 and 21 actual in-context examples for Korean, respectively. Each pair is presented at the full sentence level so the model also observes that canonical words inside a sentence should be phonemized normally. We illustrate one (user, assistant) pair per category for English in Table 13. Appendix F Downstream TTS-ASR Evaluation To examine whether token-level PER gains transfer to synthesized speech, we conduct TTS-ASR evaluations in English and Vietnamese using eSpeak NG and Whisper large-v3. System PERoverall_overall ↓ WER ↓ Canonical (gold IPA) — 0.4100.410 GPT-4o (99-shot) 0.1110.111 0.3820.382 Word ByT5 0.0780.078 0.3680.368 Ours (Qwen 2.5-0.5B) 0.1400.140 0.4360.436 Ours (ByT5) 0.0660.066 0.3590.359 Table 14: Token-level overall PER and sentence-level WER on the 100100-sentence English TTS-ASR evaluation. Lower is better. Bold: best per column. Type Can. Word GPT-4o Ours Ours ByT5 (99-shot) Qwen ByT5 A1 0.5410.541 0.4380.438 0.4810.481 0.4920.492 0.4230.423 A2 0.3960.396 0.4660.466 0.4250.425 0.5470.547 0.3790.379 A3 0.4310.431 0.4850.485 0.4620.462 0.5590.559 0.4590.459 B1 0.3630.363 0.2710.271 0.3270.327 0.3410.341 0.3030.303 B2 0.3690.369 0.3010.301 0.2660.266 0.3720.372 0.3690.369 C1 0.3920.392 0.4090.409 0.4670.467 0.4690.469 0.3780.378 C2 0.3520.352 0.3240.324 0.3560.356 0.3850.385 0.3180.318 C3 0.3870.387 0.3090.309 0.3310.331 0.3740.374 0.3130.313 C4 0.4560.456 0.3710.371 0.3610.361 0.4480.448 0.4290.429 Table 15: Sentence-level WER by UGT category on the 100100-sentence English TTS-ASR evaluation. Bold: best per row. Ours (ByT5) wins on 5/95/9 categories, with the largest gains on Reconstruct (A1–A2) and Pass-through phonetic respelling (C1–C3), where retrieved canonical-form evidence directly resolves the surface. Vietnamese evaluation. We additionally evaluate 297297 Vietnamese examples. Table 16 reports sentence-level WER with 95% bootstrap confidence intervals from 10,000 resamples. Our Qwen2.5-0.5B variant attains the lowest observed WER among predicted-input systems (0.6400.640), followed by our ByT5 variant (0.6600.660). System WER ↓ 95% CI UGT PER ↓ Gold canonical input 0.6340.634 [0.607,0.661][0.607,0.661] 0.0000.000 Word ByT5 0.6670.667 [0.640,0.694][0.640,0.694] 0.4320.432 GPT-4o (few-shot) 0.7120.712 [0.687,0.737][0.687,0.737] 0.4560.456 Ours (Qwen 2.5-0.5B) 0.6400.640 [0.613,0.666][0.613,0.666] 0.2000.200 Ours (ByT5) 0.6600.660 [0.633,0.686][0.633,0.686] 0.2720.272 Table 16: Vietnamese downstream TTS-ASR evaluation on 297297 examples. WER confidence intervals use 10,000 bootstrap resamples. Lower is better; bold marks the best predicted-input system. Setup. We sample 100100 sentences from UGTPhon’s English test split, stratified by dominant UGT macro-type (3030 Reconstruct, 2020 Repetition, 5050 Pass-through; mean length 11.711.7 tokens, 1,1691,169 valid tokens total) under seed 4242. For each system, the predicted IPA is synthesized with eSpeak NG 1.50 (en-us voice, direct IPA input) and transcribed by OpenAI Whisper large-v3 (Radford et al., 2023) (temperature 00, fp16). We compute WER between the transcript and the NeMo-normalised gold sentence. Five systems are compared: Canonical (gold IPA from GcG_c), Word ByT5 (strongest non-retrieval baseline), GPT-4o (9-shot), and our approach with the Qwen2.5-0.5B and ByT5-small backbones. Caveats. For the English evaluation, N=100N=100 makes absolute WER differences below roughly ±0.02± 0.02 unreliable without a bootstrap interval, which we do not compute. Canonical is not a strict upper bound: gold IPA contains length and secondary-stress marks (:, ”) that eSpeak NG does not always realize, so the gap to Canonical over-estimates the achievable improvement. Overall WER. Table 14 reports overall sentence-level WER. Ours (ByT5) achieves the lowest WER (0.3590.359), below Word ByT5 (0.3680.368, −0.009-0.009) and GPT-4o (0.3820.382, −0.023-0.023). The downstream ranking matches the token-level PER ranking across all four trained systems, indicating that PER on UGTPhon is a faithful proxy for downstream intelligibility on this evaluation. Ours (Qwen) underperforms (0.4360.436), consistent with its higher token-level PER (0.1400.140). Per-UGT-type WER. Table 15 breaks WER down by UGT category. Ours (ByT5) wins on 5/95/9 categories (A1, A2, C1, C2, C3), with the largest gains on recoverable shortenings: A2 (−0.087-0.087 vs. Word ByT5) and A1 (−0.015-0.015). Where our approach does not win, the failure mode is interpretable: A3 (phrasal abbreviation, e.g., idk, nvm) requires multi-word expansion that the single-token retrieval store does not cover; B-type repetition is already handled by the byte-level encoder of Word ByT5; and C4 (phrasal acronyms) benefits from GPT-4o’s broader world knowledge. Categories that benefit most from retrieval at the token level (A1–A2, C1–C3) are also where our model wins downstream, corroborating that the taxonomy localizes the same difficulty axis at both levels.