Multimodal Conversation Dataset

1Harbin Institute of Technology 2JD Explore Academy, Beijing, China


Limited by the network latency, the in-the-wild collected ViCo dataset is hardly to capture the real-time interactions between these two interlocutors. And most of the scenarios in the vico dataset are interviews, in which the "mutual" interaction between speaker and listener is not sufficient. To thoroughly modeling the conversation, we further propose the ViCo-X dataset, the recorded face-to-face conversational videos performed by two qualified actors. Compare to ViCo dataset, alongside the improved real-time, we shift the focus from listener modeling to the dyadic interaction modeling, the multi-turn dialogues are brought in as video corpus, bringing the possibility of modeling multi-turn high-level interactions while adding a modality. And the listeners' attitudes are represented by a finer-grained context-sensitive annotations: dialog act. The dialog act can not only guide the speaker's behaviours but also affect listener feedbacks. ViCo-X dataset is constructed by 10 actors, featuring 25 dialogue scenes with 26 different dialog acts. The videos are recorded in a strictly-controlled environment to provide high quality details (2k resolution) of body postures and facial expressions. About 40 minutes of conversations are recorded, annotations are accurate to 1/30 second and the average number of dialogues is 10. We propose this dataset to encourage research interests in interaction and conversation modeling, and wish it could facilitate the applications including virtual anchors, digital influencers, customer representatives, digital avatar in Metaverse.

Dataset Details

1st dataset for conversation modeling ~40 minutes duration Mandarin dialogue act

Comparison with other human conversation-related datasets

Dataset Year Public Interlocutor Multi-turn Style Environment Head motion Body motion External Anno
GRID 2006 Speaker Lab Realistic -
LRW 2016 Speaker Wild Realistic -
ObamaSet 2017 Speaker Wild Realistic -
VoxCeleb 2017 Speaker Wild Realistic -
VoxCeleb2 2018 Speaker Wild Realistic -
LRS2-BBC 2018 Speaker Wild Realistic -
LRS2-TED 2018 Speaker Wild Realistic -
Faceforensics++ 2019 Speaker Wild Realistic -
MEAD 2020 Speaker Wild Realistic emotion
Speech2Gesture 2019 Presenter Wild Realistic -
Ted Gesture 2019 Presenter Wild Realistic -
Gillies et al. 2008 Speaker, Listener Lab Simulated -
SEMAINE 2011 Speaker, Listener Lab Simulated custom dimension
Heylen et al. 2011 Speaker, Listener Lab Simulated - - -
ALICO 2014 Speaker, Listener Lab Realistic feedback signal
ViCo - Speaker, Listener Wild Realistic attitude
ViCo-X - Conversational Agent Lab Realistic dialogue act

Dialogue act distribution

dialogue act distribution
Annotated by at least three annotators with cross validation (Fleiss Kappa > 0.9)


dialogue act distribution
Ten identities in vico-x dataset


A clip of recorded video, resized for network optimization.

role role_id other_id start_time dialog_act content
A 0 3 00:00:00 initialGreeting 您好
A 0 3 00:00:22 question 请问有什么可以为您效劳的呢
Q 3 0 00:02:24 request 我要看一下自己在京东的历史消费
Q 3 0 00:06:01 question 查到了吗
Q 3 0 00:06:25 state 总额就可以
A 0 3 00:08:09 noAnswer 您好,这个总额这边是查不到的
Q 3 0 00:11:10 negativeFeedback 那你告诉我可以通过什么渠道查
Q 3 0 00:13:25 question 账单有没有
Q 3 0 00:14:20 state 账单也行
Q 3 0 00:15:15 negativeFeedback 我老婆要看,不然就说我藏私房钱
Q 3 0 00:18:17 request 你快帮我想想办法


If our dataset helps your research, please cite this website.
      title={Interactive Conversational Head Generation},
      author={Mohan Zhou and Yalong Bai and Wei Zhang and Ting Yao and Tiejun Zhao},

Ethical Use

The ViCo-X dataset would be released only for research purposes under restricted licenses.


Mohan Zhou, mhzhou99[at]outlook[dot]com