Phrase-Localized Language-Contrastive Guidance:
Training-Free Localized Accent Control for Code-Switching Text-to-Speech

Che Hyun Lee1, Sangkwon Park1, Donghun Kang1, Dongwook Lee2,
Youngho Cho2, Heeseung Kim3, Sungroh Yoon1,2,4
1Department of Electrical and Computer Engineering, Seoul National University
2Interdisciplinary Program in Artificial Intelligence, Seoul National University
3Department of Artificial Intelligence, University of Seoul
4AIIS, ASRI, INMC, and ISRC, Seoul National University

EMNLP 2026 ยท Main Conference ยท Oral Presentation

Code, the 1,200-utterance synthetic code-switching benchmark, and reproduction instructions are available at github.com/saga1214/PhraseLocalizedLCG.

Hear the difference โ€” language-learning content (ENโ†’KO)

Today's key expression is ์‹œ๊ฐ„ ๋‚ด ์ฃผ์…”์„œ ์ •๋ง ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค, which you can use at the end of a meeting.
VanillaListen to the Korean phrase with the English carrier's language conditioning.
Ours โ€” LCGCompare the same phrase with localized Korean-language guidance.

What is LCG?

  1. The problem. When a sentence embeds a foreign phrase ("This station is ํ™๋Œ€์ž…๊ตฌ"), zero-shot TTS speaks the phrase with the carrier language's accent โ€” because classifier-free guidance applies one language condition to the whole utterance.
  2. The idea. Add a third guidance branch conditioned on the phrase's own language, and apply it only inside the phrase region.
  3. The trick. The phrase region is located by the model itself: we read its self-attention to find which audio frames align to the embedded text โ€” no fine-tuning, no external aligner.

Overview

The challenge

Global guidance flattens the embedded phrase into the carrier accent (left); LCG steers only the phrase region toward its own language (right).

Cross-lingual accent leakage and localized control

The method in three steps

  1. Probe โ€” read decoder self-attention (layers 8 & 12) to find the frames aligned to the embedded phrase.
  2. Refine โ€” union the masks from both language tags, then dilate for recall; the diffusion process absorbs the slack.
  3. Guide โ€” add a phrase-language guidance term, gated by the mask, with its own strength ฮป.
Method overview: attention probing, mask refinement, phrase-localized LCG

Real-World Scenarios

Everyday situations where a carrier sentence embeds a foreign phrase. Vanilla vs. our phrase-localized LCG (ฮป=7), paper-final configuration.

The last tab embeds French and German phrases in English carriers โ€” the training-free control covers all five languages of the paper.

Language-learning content ENโ†’KO

Today's key expression is ์‹œ๊ฐ„ ๋‚ด ์ฃผ์…”์„œ ์ •๋ง ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค, which you can use at the end of a meeting.
ยป ์‹œ๊ฐ„ ๋‚ด ์ฃผ์…”์„œ ์ •๋ง ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค โ€” thank you so much for making the time
Vanilla
Ours ฮป7

Tour guide ENโ†’KO

Our next stop is ์กฐ์„  ์™•์กฐ์˜ ์ •๊ถ์ธ ๊ฒฝ๋ณต๊ถ, followed by a walk through ๋ถ์ดŒ ํ•œ์˜ฅ๋งˆ์„์˜ ์ข์€ ๊ณจ๋ชฉ๊ธธ.
ยป ์กฐ์„  ์™•์กฐ์˜ ์ •๊ถ์ธ ๊ฒฝ๋ณต๊ถ โ€” Gyeongbokgung, the main royal palace of the Joseon dynasty
ยป ๋ถ์ดŒ ํ•œ์˜ฅ๋งˆ์„์˜ ์ข์€ ๊ณจ๋ชฉ๊ธธ โ€” the narrow alleys of Bukchon Hanok Village
Vanilla
Ours ฮป7

Museum exhibition ENโ†’JA

The exhibition features ๆฑŸๆˆธๆ™‚ไปฃใฎๆตฎไธ–็ตตใฎๅไฝœ from the museum's own collection.
ยป ๆฑŸๆˆธๆ™‚ไปฃใฎๆตฎไธ–็ตตใฎๅไฝœ โ€” masterpieces of Edo-period ukiyo-e woodblock prints
Vanilla
Ours ฮป7

Benchmark Samples

Samples drawn directly from our released 1,200-utterance synthetic code-switching benchmark, comparing all three systems from the paper (unguided baseline, coupled Swap, and phrase-localized LCG).

Isolated phrase segments

Korean carrier sentences with embedded English phrases, drawn from the LA=1 subset: all systems synthesize the correct words, so the difference you hear is residual accent quality on the isolated segment.

Example 1 โ€” โ€œpractical workshops aligned with our buying committeeโ€ KOโ†’EN

์ง€๋‚œ๋‹ฌ ์šฐ๋ฆฌ ํŒ€์ด ๋„์ž…ํ•œ ๊ธฐ์—… ๊ฐ„ ์˜์—… ์ปจ์„คํŒ…์‚ฌ๋Š” ๊ธฐ๋Œ€๋ณด๋‹ค ํ›จ์”ฌ ์‹ค๋ฌด์ ์ด์—ˆ๊ณ , ๋ฆฌ๋ทฐ๋ฅผ ๋‚จ๊ธธ ๋งŒํผ ์ธ์ƒ์ด ์ข‹์•˜๋‹ค. ์ฒซ ์ง„๋‹จ ๋ฏธํŒ…์—์„œ a clear account-based sales playbook ์œผ๋กœ ๋…ผ์˜๋ฅผ ์ •๋ฆฌํ•ด ์ฃผ๋‹ˆ, ์˜์—…๋Œ€ํ‘œ์™€ ๋งˆ์ผ€ํŒ… ๋‹ด๋‹น์ž๊ฐ€ ๊ฐ™์€ ์–ธ์–ด๋กœ ๋ฌธ์ œ๋ฅผ ๋ณด๊ฒŒ ๋๋‹ค. ์›Œํฌ์ˆ์€ practical workshops aligned with our buying committee ์™€ ์‹ค์ œ ๊ณ ๊ฐ ๋ฏธํŒ… ์‚ฌ๋ก€๋ฅผ ์—ฎ์–ด ์ง„ํ–‰๋ผ, ๊ฐ•์˜๋ณด๋‹ค ๊ณต๋™ ์„ค๊ณ„์— ๊ฐ€๊นŒ์› ๋‹ค. ํŠนํžˆ actionable pipeline diagnostics grounded in real CRM data ๊ฐ€ ๊ฐ€์žฅ ๋„์›€์ด ๋๋Š”๋ฐ, ๋ง‰์—ฐํ–ˆ๋˜ ๋ฆฌ๋“œ ํ’ˆ์งˆ ๋ฌธ์ œ์™€ ๋”œ ์ง€์—ฐ ์›์ธ์ด ์ˆซ์ž๋กœ ๋ณด์˜€๋‹ค. ๋งˆ์ง€๋ง‰ ์ œ์•ˆ์„œ์— ๋‹ด๊ธด a lightweight operating model for cross-functional execution ๊ณผ transparent ROI assumptions for the first six months ๊นŒ์ง€ ์„ค๋“๋ ฅ ์žˆ์–ด, ๋น„์šฉ์€ ๋†’์ง€๋งŒ ์žฌ๊ณ„์•ฝ์€ ๊ธ์ •์ ์œผ๋กœ ๊ฒ€ํ†  ์ค‘์ด๋‹ค.
Vanilla
Swap ฮป3
Ours ฮป7

Example 2 โ€” โ€œlean household operations with zero emotional overheadโ€ KOโ†’EN

ํ‡ด๊ทผํ•˜๊ณ  ์ง‘์— ์˜ค์ž๋งˆ์ž ์šด๋™๋ณต์„ ์„ธํƒ๊ธฐ์— ๋„ฃ์—ˆ๋Š”๋ฐ, ์˜ค๋Š˜์€ minimum viable weekday morning routine ์œผ๋กœ ๋‚ด์ผ ์ž…์„ ์–‘๋ง์ด๋ž‘ ๋จธ๊ทธ์ปต๊นŒ์ง€ ๋ฏธ๋ฆฌ ์ฑ™๊ฒผ๋‹ค. ์‹ฑํฌ๋Œ€์— ์ปต ๋‘ ๊ฐœ๋ž‘ ์ ‘์‹œ ํ•˜๋‚˜๋งŒ ๋‚จ๊ฒจ๋‘๊ธฐ ์‹ซ์–ด์„œ lean household operations with zero emotional overhead ์ฒ˜๋Ÿผ ๋ฐ”๋กœ ์น˜์› ๊ณ , ๋ƒ‰์žฅ๊ณ  ์•ž ๋ฉ”๋ชจ์—๋Š” data driven grocery budget tracking system ์˜ ์ž”์•ก์„ ์ž‘๊ฒŒ ์ ์–ด๋†จ๋‹ค. ๋ฒ ๋ž€๋‹ค ํ™”๋ถ„์— ๋ฌผ ์ฃผ๋Š” ์‚ฌ์ด ๊ณ ์–‘์ด๋Š” ์ƒˆ ํƒ๋ฐฐ ๋ฐ•์Šค์— ๋“ค์–ด๊ฐ€์„œ weekly user research session with a very picky stakeholder ์— ์ฐธ์—ฌํ•˜๋Š” ํ‘œ์ •์œผ๋กœ ๋‚˜๋ฅผ ๋ดค๋‹ค. ๋ณ„๊ฑฐ ์•„๋‹Œ ์ง‘์•ˆ์ผ๋„ ์ด๋ ‡๊ฒŒ ์ด๋ฆ„์„ ๋ถ™์ด๋ฉด ๋ฌ˜ํ•˜๊ฒŒ ๊ฒŒ์ž„ ๊ฐ™๊ณ , ์ž ๋“ค๊ธฐ ์ „ ๋งˆ์Œ์ด ์กฐ๊ธˆ ์ •๋ฆฌ๋œ๋‹ค.
Vanilla
Swap ฮป3
Ours ฮป7

Full sentences

Complete utterances with three embedded phrases each, comparing global naturalness and phrase nativeness.

English carrier ยท Japanese phrases ENโ†’JA

After the rain, I ducked into a tiny old-school cafe by the station, and the ๆทฑ็…Žใ‚Š่ฑ†ใ‚’ใƒใƒซใง่ฝใจใ™้ฆ™ใ‚Š made my whole jacket smell like comfort. The owner gave me a smile that said ใ”ใ‚†ใฃใใ‚ŠใŠๅฌใ—ไธŠใŒใ‚Šใใ ใ•ใ„, not fussy, just warm. I took the corner seat and watched steam blur the window, feeling a sort of ้›จไธŠใŒใ‚Šใฎ็ธๅดใง้ฃฒใ‚€ใปใ†ใ˜่Œถ settle into the evening. Next time, Iโ€™m bringing a book and pretending I donโ€™t have notifications. Honestly, it beat any fancy latte art.
ยป ๆทฑ็…Žใ‚Š่ฑ†ใ‚’ใƒใƒซใง่ฝใจใ™้ฆ™ใ‚Š โ€” the aroma of dark-roasted beans brewed through a flannel filter
ยป ใ”ใ‚†ใฃใใ‚ŠใŠๅฌใ—ไธŠใŒใ‚Šใใ ใ•ใ„ โ€” please take your time and enjoy it
ยป ้›จไธŠใŒใ‚Šใฎ็ธๅดใง้ฃฒใ‚€ใปใ†ใ˜่Œถ โ€” roasted green tea sipped on a veranda after the rain
Vanilla
Swap ฮป3
Ours ฮป7

English carrier ยท Korean phrases ENโ†’KO

Hanโ€™s slim novel reads like a review of everything we keep misplacing: names, rooms, the voices of the dead. The ์ฒซ๋ˆˆ ๋…น๋“ฏ ์‚ฌ๋ผ์ง€๋Š” ์–ด๋จธ๋‹ˆ์˜ ๋งˆ์ง€๋ง‰ ์†๊ธธ gives the opening its ache, while the ์ œ์‚ฌ์ƒ์— ์˜ฌ๋ฆฐ ์‹์–ด ๊ฐ€๋Š” ํฐ๋ฐฅ ํ•œ ๊ทธ๋ฆ‡ turns grief into something almost domestic. By the final chapter, even ๋ด„๋น„์— ์ –์€ ๋ฒš๊ฝƒ์žŽ์ด ํฉ์–ด์ง„ ๊ณจ๋ชฉ feels less symbolic than inevitable. I admired its restraint; it left me quiet, not devastated, and somehow that was worse.
ยป ์ฒซ๋ˆˆ ๋…น๋“ฏ ์‚ฌ๋ผ์ง€๋Š” ์–ด๋จธ๋‹ˆ์˜ ๋งˆ์ง€๋ง‰ ์†๊ธธ โ€” last trace of a motherโ€™s touch fading like first snow
ยป ์ œ์‚ฌ์ƒ์— ์˜ฌ๋ฆฐ ์‹์–ด ๊ฐ€๋Š” ํฐ๋ฐฅ ํ•œ ๊ทธ๋ฆ‡ โ€” cooling bowl of white rice set out for the ancestors
ยป ๋ด„๋น„์— ์ –์€ ๋ฒš๊ฝƒ์žŽ์ด ํฉ์–ด์ง„ ๊ณจ๋ชฉ โ€” an alley strewn with cherry petals wet from spring rain
Vanilla
Swap ฮป3
Ours ฮป7

Evaluation Corpus

All audio above is synthesized from our 1,200-utterance balanced synthetic code-switching benchmark, generated with a large language model and spanning five languages across twelve directional configurations. The corpus is released with the code and documented in benchmark/.

1,200
utterances
12
directions
5
languages (EN/FR/DE/JA/KO)
~3.6
avg phrases/utt
18
topical domains
6
stylistic registers

Each utterance contains 3โ€“5 dense, technical or literary embedded phrases (~6 words, ~35 characters per phrase). The corpus is balanced by direction (100 utterances per direction) to enable per-direction evaluation.

Citation

@misc{lee2026phraselocalizedlanguagecontrastiveguidancetrainingfree,
  title         = {Phrase-Localized Language-Contrastive Guidance: Training-Free Localized Accent Control for Code-Switching Text-to-Speech},
  author        = {Che Hyun Lee and Sangkwon Park and Donghun Kang and Dongwook Lee and Youngho Cho and Heeseung Kim and Sungroh Yoon},
  year          = {2026},
  eprint        = {2609.01016},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.01016},
}