# ArXiv Preprint Introduces UGTPhon, a Benchmark for Pronouncing Slang and Abbreviations

> An arXiv preprint from ArXiv CS.CL introduces UGTPhon, a G2P benchmark showing existing models and frontier LLMs miss user-generated English, Vietnamese, and Korean text by up to 66.8 PER points.

- **Topic**: Models
- **Published**: 2026-09-24T07:49:07.725Z
- **Canonical URL**: https://highsignal.sh/stories/arxiv-preprint-introduces-ugtphon-a-benchmark-for-pronouncing-slang-and-abbrevia-15207607

## Why It Matters

Text-to-speech systems increasingly read user-generated text, where spellings like 'ppl' and 'imo' don't sound like they read. A measured gap of up to 66.8 PER points suggests current phonemization methods fail systematically on exactly the input real users type.

## Key Findings & Analysis

### What the preprint reports

An arXiv paper (authors not named in the supplied abstract) introduces UGTPhon, described as the first grapheme-to-phoneme (G2P) benchmark for user-generated text in English, Vietnamese, and Korean, plus an inference-grounded taxonomy for fine-grained diagnosis. The authors say existing G2P models and frontier LLMs show a systematic canonical-to-non-canonical performance gap reaching up to 66.8 PER points. As a baseline, they propose a compositional G2P approach that uses canonical-form evidence through exact-match lookup and staged decoding. Across matched ByT5 and Qwen2.5-0.5B backbones, the paper reports explicit canonical-form modeling consistently reduces non-canonical G2P errors, and that the 0.5B variant is competitive with much larger few-shot frontier LLMs. All figures and comparisons here are the authors' own claims; the abstract provides no independent replication.

## Primary Sources & Citations

- [Phonemizing User-Generated Text: A Benchmark, Taxonomy, and Compositional Approach](https://arxiv.org/abs/2609.27205) — *ArXiv CS.CL (Computation and Language)* (Reporting)

---

[← Back to front page](https://highsignal.sh/) | [Daily Brief](https://highsignal.sh/brief) | [All stories](https://highsignal.sh/latest)