Automated Video Creation from Audio

With the growing reliance on video in distance learning applications, the costs of producing such video can become an impediment for distance learning producers. To address this issue Byung-Hak Kim and Varun Ganapathi, researchers at Udacity, have developed LumiereNet, which they describe as “a simple, modular, and completely deep-learning based architecture that synthesizes high quality, full-pose headshot lecture videos from instructor’s new audio narration of any length.” The system is described in their paper LumiereNet: Lecture Video Synthesis from Audio.

The paper steps through the process of developing the system and includes two video clips of the resulting video products:

https://vimeo.com/327196781
https://vimeo.com/327196551

The results are encouraging as shown in the video clips and the researchers identify areas for further development and refinement of the process to achieve better results moving forward.

There are several practical applications of this process. For example, with several hours of video footage of an instructor, it is possible to automatically create updated video clips when an instructor updates material in an online course. This will reduce the calls on the instructor’s time as well as the cost of additional video production. As the automated process improves, this option for updating online courses with video components will become increasingly attractive. In addition, it may be possible to automate corrections to video footage even when an instructor is only engaged for a single run-through of a particular lesson.

One issue not addressed by the researchers is the learning value of instructor headshot video. Although many online offerings include headshot video of instructors, the value of such video for learning is not clear and may be worth investigation. Still video may help establish a sense of instructor presence and connection to students even if there is little direct contribution to learning particular content.

Be the first to comment

Leave a Reply