@inproceedings{nguyen-etal-2026-tts2,
title = "{TTS}$^2$: Test-Time Scaling for Flow-Based Text-to-Speech via Multi-stage Reward Modeling on Posterior Mean Projections",
author = "Nguyen, Tung Luong and
Le, Tuan Duy Nguyen and
Tran, Nhat Hong and
Le, Nhat Minh and
Do, Van Hai",
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.2541/",
pages = "57786--57797",
ISBN = "979-8-89176-459-0",
abstract = "While training-time scaling has revolutionized generative models, inference-time compute scaling remains largely unexplored in speech synthesis. In this work, we investigate methods to effectively utilize computation during inference to improve the performance of non-autoregressive Text-to-Speech (TTS) systems, specifically those based on Conditional Flow Matching (CFM). By analyzing CFM synthesis behavior, we identify distinct performance plateaus during the denoising process. To overcome these limitations, we propose TTS{\texttwosuperior} - a multi-stage, test-time compute framework that optimizes TTS quality. By combining Ordinary Differential Equation (ODE) and Stochastic Differential Equation (SDE) generation dynamics with stochastic search and specialized verifiers, our approach enables guided trajectory optimization for timbre consistency, content accuracy, and audio fidelity. To the best of our knowledge, this work is the first to formalize and study test-time scaling for non-autoregressive speech synthesis. Source code, models, and audio demos are publicly available at https://anonymous.4open.science/r/TTS2-Test-Time-Scaling-for-Flow-Based-Text-to-Speech-8645."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="nguyen-etal-2026-tts2">
<titleInfo>
<title>TTS²: Test-Time Scaling for Flow-Based Text-to-Speech via Multi-stage Reward Modeling on Posterior Mean Projections</title>
</titleInfo>
<name type="personal">
<namePart type="given">Tung</namePart>
<namePart type="given">Luong</namePart>
<namePart type="family">Nguyen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tuan</namePart>
<namePart type="given">Duy</namePart>
<namePart type="given">Nguyen</namePart>
<namePart type="family">Le</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nhat</namePart>
<namePart type="given">Hong</namePart>
<namePart type="family">Tran</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nhat</namePart>
<namePart type="given">Minh</namePart>
<namePart type="family">Le</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Van</namePart>
<namePart type="given">Hai</namePart>
<namePart type="family">Do</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>While training-time scaling has revolutionized generative models, inference-time compute scaling remains largely unexplored in speech synthesis. In this work, we investigate methods to effectively utilize computation during inference to improve the performance of non-autoregressive Text-to-Speech (TTS) systems, specifically those based on Conditional Flow Matching (CFM). By analyzing CFM synthesis behavior, we identify distinct performance plateaus during the denoising process. To overcome these limitations, we propose TTS² - a multi-stage, test-time compute framework that optimizes TTS quality. By combining Ordinary Differential Equation (ODE) and Stochastic Differential Equation (SDE) generation dynamics with stochastic search and specialized verifiers, our approach enables guided trajectory optimization for timbre consistency, content accuracy, and audio fidelity. To the best of our knowledge, this work is the first to formalize and study test-time scaling for non-autoregressive speech synthesis. Source code, models, and audio demos are publicly available at https://anonymous.4open.science/r/TTS2-Test-Time-Scaling-for-Flow-Based-Text-to-Speech-8645.</abstract>
<identifier type="citekey">nguyen-etal-2026-tts2</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2541/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>57786</start>
<end>57797</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T TTS²: Test-Time Scaling for Flow-Based Text-to-Speech via Multi-stage Reward Modeling on Posterior Mean Projections
%A Nguyen, Tung Luong
%A Le, Tuan Duy Nguyen
%A Tran, Nhat Hong
%A Le, Nhat Minh
%A Do, Van Hai
%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 nguyen-etal-2026-tts2
%X While training-time scaling has revolutionized generative models, inference-time compute scaling remains largely unexplored in speech synthesis. In this work, we investigate methods to effectively utilize computation during inference to improve the performance of non-autoregressive Text-to-Speech (TTS) systems, specifically those based on Conditional Flow Matching (CFM). By analyzing CFM synthesis behavior, we identify distinct performance plateaus during the denoising process. To overcome these limitations, we propose TTS² - a multi-stage, test-time compute framework that optimizes TTS quality. By combining Ordinary Differential Equation (ODE) and Stochastic Differential Equation (SDE) generation dynamics with stochastic search and specialized verifiers, our approach enables guided trajectory optimization for timbre consistency, content accuracy, and audio fidelity. To the best of our knowledge, this work is the first to formalize and study test-time scaling for non-autoregressive speech synthesis. Source code, models, and audio demos are publicly available at https://anonymous.4open.science/r/TTS2-Test-Time-Scaling-for-Flow-Based-Text-to-Speech-8645.
%U https://aclanthology.org/2026.emnlp-main.2541/
%P 57786-57797
Markdown (Informal)
[TTS2: Test-Time Scaling for Flow-Based Text-to-Speech via Multi-stage Reward Modeling on Posterior Mean Projections](https://aclanthology.org/2026.emnlp-main.2541/) (Nguyen et al., EMNLP 2026)
ACL