Online meetings have become a frequently used tool for business and learning. How to meet the more diversifying online conferencing needs of users has brought great challenges to remote work applications, including captioning, real-time machine translation, smart meeting minutes and other artificial intelligence applications.

The core technology of smart web conferencing is ASR (Automatic Speech Recognition). For conference audio, the biggest challenge is how to continuously detect and recognize speaker switching in multi-person conference scenes. Machine learning training on big data is what will help make this technology a reality.

However, licensable multi-person meeting data in real scenarios is limited. For this, specially releases a dataset of multi-person spontaneous meetings, including 400 meetings with a total of 200 hours. The topics of the meetings include company activities, employee benefits, employee training, product design, product optimization, marketing activities, business management, team management, and family life. etc., which are all conference content based on real business and life scenarios. No script is provided, the speakers express themselves spontaneously.

This data set is collected in conference rooms, subways, parks, airports, shopping malls, railway stations, airports and other scenes. More than 400 people participated in the collection. In addition to the built-in recording of mainstream conference software on the market, each speaker is also equipped with a separate recording device for the data collection, including headset, computers, recorder, mobile phones, tablets, etc. This can meet the research and development needs of speech recognition and speaker recognition under multi-channel and single-channel.

Data Specification

Transcription of the conference audio is attached in this dataset. More than that, for the research and development needs of smart meeting minutes, has conducted natural language processing annotation on the transcription, extracting key conference information such as speakers, participants, topics, key points, conference conclusions, conference decisions, and to-do list. Consecutive speeches by the same person in the transcription are merged into one segment, which prevents the labeling result from being excessively split into multiple fragments. The tagged content in the transcription supports index, which can be traced back to the original text by character. An example of labeling is as follows:

[{"start": 3, "end": 19, "text": "We will hold a team building decision-making meeting online today", "labels": ["theme"]},{""start"": 20 , ""end"": 47, ""text"": ""Then this meeting is chaired by Sister Li, and the participants include Feifei, Chen Chen, and Xiaonan"", ""labels"": ["" participant""]}]" [{"start": 116, "end": 133, "text": "Then Chen Chen and Feifei will talk about their respective plans", "labels": ["speaker"]}]"

hopes that this Multi-Person Spontaneous Meetings Dataset can help the web conference R&D team quickly realize product iterations, build leading smart conference system, and win more market opportunities.

Contact business@magicdatatech.com to learn more details and obtain a total smart web conferencing data solution.