@inproceedings{li-etal-2026-unlocking-spoken,
title = "Unlocking the Spoken Style: Mechanistic Localization and Lightweight Control of Orality in {LLM}s",
author = "Li, Zhigen and
Wang, Yanmeng and
Peng, Jianxiang and
Zhang, Yong and
Shi, Dan and
Cheng, Ning and
Wang, Shaojun and
Xiong, Deyi",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.381/",
pages = "8706--8726",
ISBN = "979-8-89176-459-0",
abstract = "Voice agents typically use text-first pipeline for safety, but large language model (LLM) outputs are often overly written, making even high-quality text-to-speech (TTS) sound robotic. This is especially problematic in domains such as healthcare and sales, where different audiences require graded orality control. Text-level rewriting can control spoken style, but resources mostly study the reverse direction (spoken-to-written transfer), and its internal locus in LLMs remains unclear. We study English written-to-spoken rewriting with graded orality control. We introduce WSPairs, 4{,}709 written{--}spoken pairs from spontaneous conversations, annotated with 0{--}5 Orality Levels using seven sociolinguistic features. Through causal activation patching and head-level attribution, we find that orality concentrates in late layers and a sparse set of high-influence attention heads. We then propose CAS, a no-weight-update calibrated activation-steering method, and LAS, a learned activation-steering method. CAS improves exact level targeting and reduces bias relative to few-shot prompting, although target ordering remains imperfect at the highest levels. LAS uses 0.004{\%}{--}0.006{\%} trainable parameters, obtains higher Orality Level Accuracy (L-Acc) than full SFT on Qwen-3B and Mistral-7B, remains the strongest lightweight method on Llama-1B, and leads SFT by up to 3.9 L-Acc points out of domain. These results support localized, efficient control of text-level spoken style. Dataset and code are available."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-etal-2026-unlocking-spoken">
<titleInfo>
<title>Unlocking the Spoken Style: Mechanistic Localization and Lightweight Control of Orality in LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zhigen</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yanmeng</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jianxiang</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yong</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dan</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ning</namePart>
<namePart type="family">Cheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shaojun</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Deyi</namePart>
<namePart type="family">Xiong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Voice agents typically use text-first pipeline for safety, but large language model (LLM) outputs are often overly written, making even high-quality text-to-speech (TTS) sound robotic. This is especially problematic in domains such as healthcare and sales, where different audiences require graded orality control. Text-level rewriting can control spoken style, but resources mostly study the reverse direction (spoken-to-written transfer), and its internal locus in LLMs remains unclear. We study English written-to-spoken rewriting with graded orality control. We introduce WSPairs, 4,709 written–spoken pairs from spontaneous conversations, annotated with 0–5 Orality Levels using seven sociolinguistic features. Through causal activation patching and head-level attribution, we find that orality concentrates in late layers and a sparse set of high-influence attention heads. We then propose CAS, a no-weight-update calibrated activation-steering method, and LAS, a learned activation-steering method. CAS improves exact level targeting and reduces bias relative to few-shot prompting, although target ordering remains imperfect at the highest levels. LAS uses 0.004%–0.006% trainable parameters, obtains higher Orality Level Accuracy (L-Acc) than full SFT on Qwen-3B and Mistral-7B, remains the strongest lightweight method on Llama-1B, and leads SFT by up to 3.9 L-Acc points out of domain. These results support localized, efficient control of text-level spoken style. Dataset and code are available.</abstract>
<identifier type="citekey">li-etal-2026-unlocking-spoken</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.381/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>8706</start>
<end>8726</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Unlocking the Spoken Style: Mechanistic Localization and Lightweight Control of Orality in LLMs
%A Li, Zhigen
%A Wang, Yanmeng
%A Peng, Jianxiang
%A Zhang, Yong
%A Shi, Dan
%A Cheng, Ning
%A Wang, Shaojun
%A Xiong, Deyi
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F li-etal-2026-unlocking-spoken
%X Voice agents typically use text-first pipeline for safety, but large language model (LLM) outputs are often overly written, making even high-quality text-to-speech (TTS) sound robotic. This is especially problematic in domains such as healthcare and sales, where different audiences require graded orality control. Text-level rewriting can control spoken style, but resources mostly study the reverse direction (spoken-to-written transfer), and its internal locus in LLMs remains unclear. We study English written-to-spoken rewriting with graded orality control. We introduce WSPairs, 4,709 written–spoken pairs from spontaneous conversations, annotated with 0–5 Orality Levels using seven sociolinguistic features. Through causal activation patching and head-level attribution, we find that orality concentrates in late layers and a sparse set of high-influence attention heads. We then propose CAS, a no-weight-update calibrated activation-steering method, and LAS, a learned activation-steering method. CAS improves exact level targeting and reduces bias relative to few-shot prompting, although target ordering remains imperfect at the highest levels. LAS uses 0.004%–0.006% trainable parameters, obtains higher Orality Level Accuracy (L-Acc) than full SFT on Qwen-3B and Mistral-7B, remains the strongest lightweight method on Llama-1B, and leads SFT by up to 3.9 L-Acc points out of domain. These results support localized, efficient control of text-level spoken style. Dataset and code are available.
%U https://aclanthology.org/2026.emnlp-main.381/
%P 8706-8726
Markdown (Informal)
[Unlocking the Spoken Style: Mechanistic Localization and Lightweight Control of Orality in LLMs](https://aclanthology.org/2026.emnlp-main.381/) (Li et al., EMNLP 2026)
ACL
- Zhigen Li, Yanmeng Wang, Jianxiang Peng, Yong Zhang, Dan Shi, Ning Cheng, Shaojun Wang, and Deyi Xiong. 2026. Unlocking the Spoken Style: Mechanistic Localization and Lightweight Control of Orality in LLMs. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 8706–8726, Budapest, Hungary. Association for Computational Linguistics.