@inproceedings{zhang-etal-2026-embodichat,
title = "{E}mbodi{C}hat: Towards Zero-Latency Text-driven Talking Face Generation for Embodied Dialogue",
author = "Zhang, Tiecheng and
Yu, Huan and
Runjie, Zhu and
Xu, Zequan and
Wang, Wei and
Huang, Kaixiang and
Wang, Jin",
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.267/",
pages = "6070--6095",
ISBN = "979-8-89176-459-0",
abstract = "Seamless translation of natural language into embodied behaviors remains critical for open domain embodied dialogue. We formalize this task as mapping large language model text to synchronized speech and facial animations. Prevailing cascaded Text to Speech (TTS) and Talking Face Generation (TFG) pipelines suffer from prohibitive latency and poor cross platform generalizability. We present EmbodiChat, an end-to-end real-time dialogue system overcoming these bottlenecks. The key insight of EmbodiChat is that TTS acoustic representations capture pure ``articulatory intent,'' making them far more robust than audio-derived semantic features. Aligning facial generation directly onto this representation without additional adapters enables decoupled audio visual synthesis, fundamentally eliminating the serial latency of conventional audio first architectures. Concurrently, our framework decodes these representations into hardware agnostic action sequences. EmbodiChat achieves a stable Time-to-First-Audio (TTFA) of about 150 ms on a consumer GPU, significantly outperforming mainstream methods. Furthermore, it delivers a 3x speedup relative to its ablated audio driven baseline while maintaining comparable animation quality. Deployment across two distinct embodied agent platforms configurations demonstrates a robust, practically deployable text driven solution for embodied dialogue. The data, code and model will be released at https://github.com/JiangMingZi/Embodi{\_}Chat"
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhang-etal-2026-embodichat">
<titleInfo>
<title>EmbodiChat: Towards Zero-Latency Text-driven Talking Face Generation for Embodied Dialogue</title>
</titleInfo>
<name type="personal">
<namePart type="given">Tiecheng</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Huan</namePart>
<namePart type="family">Yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhu</namePart>
<namePart type="family">Runjie</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zequan</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kaixiang</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jin</namePart>
<namePart type="family">Wang</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>Seamless translation of natural language into embodied behaviors remains critical for open domain embodied dialogue. We formalize this task as mapping large language model text to synchronized speech and facial animations. Prevailing cascaded Text to Speech (TTS) and Talking Face Generation (TFG) pipelines suffer from prohibitive latency and poor cross platform generalizability. We present EmbodiChat, an end-to-end real-time dialogue system overcoming these bottlenecks. The key insight of EmbodiChat is that TTS acoustic representations capture pure “articulatory intent,” making them far more robust than audio-derived semantic features. Aligning facial generation directly onto this representation without additional adapters enables decoupled audio visual synthesis, fundamentally eliminating the serial latency of conventional audio first architectures. Concurrently, our framework decodes these representations into hardware agnostic action sequences. EmbodiChat achieves a stable Time-to-First-Audio (TTFA) of about 150 ms on a consumer GPU, significantly outperforming mainstream methods. Furthermore, it delivers a 3x speedup relative to its ablated audio driven baseline while maintaining comparable animation quality. Deployment across two distinct embodied agent platforms configurations demonstrates a robust, practically deployable text driven solution for embodied dialogue. The data, code and model will be released at https://github.com/JiangMingZi/Embodi_Chat</abstract>
<identifier type="citekey">zhang-etal-2026-embodichat</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.267/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>6070</start>
<end>6095</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T EmbodiChat: Towards Zero-Latency Text-driven Talking Face Generation for Embodied Dialogue
%A Zhang, Tiecheng
%A Yu, Huan
%A Runjie, Zhu
%A Xu, Zequan
%A Wang, Wei
%A Huang, Kaixiang
%A Wang, Jin
%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 zhang-etal-2026-embodichat
%X Seamless translation of natural language into embodied behaviors remains critical for open domain embodied dialogue. We formalize this task as mapping large language model text to synchronized speech and facial animations. Prevailing cascaded Text to Speech (TTS) and Talking Face Generation (TFG) pipelines suffer from prohibitive latency and poor cross platform generalizability. We present EmbodiChat, an end-to-end real-time dialogue system overcoming these bottlenecks. The key insight of EmbodiChat is that TTS acoustic representations capture pure “articulatory intent,” making them far more robust than audio-derived semantic features. Aligning facial generation directly onto this representation without additional adapters enables decoupled audio visual synthesis, fundamentally eliminating the serial latency of conventional audio first architectures. Concurrently, our framework decodes these representations into hardware agnostic action sequences. EmbodiChat achieves a stable Time-to-First-Audio (TTFA) of about 150 ms on a consumer GPU, significantly outperforming mainstream methods. Furthermore, it delivers a 3x speedup relative to its ablated audio driven baseline while maintaining comparable animation quality. Deployment across two distinct embodied agent platforms configurations demonstrates a robust, practically deployable text driven solution for embodied dialogue. The data, code and model will be released at https://github.com/JiangMingZi/Embodi_Chat
%U https://aclanthology.org/2026.emnlp-main.267/
%P 6070-6095
Markdown (Informal)
[EmbodiChat: Towards Zero-Latency Text-driven Talking Face Generation for Embodied Dialogue](https://aclanthology.org/2026.emnlp-main.267/) (Zhang et al., EMNLP 2026)
ACL